Energy per Successful Goal: Goal-Level Energy Accounting for Agentic AI Systems

# Goal-Level Energy Accounting for Agentic AI Systems
Source: [https://arxiv.org/html/2605.22883](https://arxiv.org/html/2605.22883)
andAakash TyagiTexas A&M UniversityDepartment of Computer Science and EngineeringUSA[tyagi@cse\.tamu\.edu](https://arxiv.org/html/2605.22883v1/mailto:[email protected])

###### Abstract\.

Current AI energy benchmarks measure consumption at the granularity of a single model invocation or training run\. For classical single\-turn workloads this unit remains coherent\. For agentic systems — where a single user goal may trigger multi\-step orchestration, tool calls, retries, and failure\-recovery cycles — the invocation count is an implementation artifact rather than a task property, and inference\-level normalization systematically misrepresents the true energy cost of goal completion\.

We presentA\-LEMS\(Agentic LLM Energy Measurement System\), a cross\-layer measurement framework that redefines the fundamental unit of AI energy accounting from*energy per inference*toEnergy per Successful Goal\(EpG\)\.EpGaggregates total workflow energy across all execution attempts, including failures and retries, normalized by successfully completed goals\. This shifts energy evaluation from low\-level model calls to end\-to\-end task completion, aligning measurement with how agentic systems are actually used\.

Beyond redefining the unit,A\-LEMSformalizes energy attribution through three coupled components: a temporal boundary model that defines the attribution window of a workflow; a five\-layer observation pipeline that maps hardware\-levelRAPLsignals through baseline subtraction and CPU\-fraction attribution to workflow\-level energy; and a reproducibility protocol that binds every measurement to hardware, environment, and runtime configuration\. Building onEpG, we define the Orchestration Overhead Index \(OOI\), which isolates the additional energy cost induced by multi\-step orchestration relative to linear execution on the same goal instances under identical task and success criteria, enabling consistent measurement across agentic and non\-agentic systems\.

Across five reasoning task families \(factual QA, science QA, arithmetic reasoning, multi\-step reasoning, and logical reasoning\) and three tool\-augmented task families with hardware\-levelRAPLenergy profiling, we find that agentic workflows consume4\.33×4\.33\\timeshigher mean energy per successful goal compared to equivalent linear baselines \(888\.1888\.1J vs205\.3205\.3J\)\. This overhead is not driven by increased inference compute but by orchestration structure, including retries, intermediate planning, and recovery behavior\. Critically, for tool\-augmented tasks where agentic dispatch replaces costly token generation,OOIinverts below1\.0×1\.0\\times: agentic execution is strictly cheaper than linear, confirming that the metric responds to orchestration structure rather than imposing a fixed upward bias\.

These findings establish that energy\-per\-inference is insufficient for evaluating modern AI systems: goal\-level accounting viaEpGandOOIprovides the measurement foundation for accurate benchmarking and system design in agentic AI workloads, where orchestration structure, not inference substrate, is the primary determinant of energy cost\.

energy measurement, agentic AI, RAPL, EpG, orchestration overhead, LLM benchmarking, reproducibility, green AI

††ccs:Computing methodologies Artificial intelligence††ccs:Computing methodologies Multi\-agent systems††ccs:Hardware Power and energy††ccs:Computer systems organization Energy\-aware systems††ccs:Software and its engineering Software performance††ccs:General and reference Measurement††ccs:Software and its engineering Software verification and validation††ccs:Software and its engineering Software notations and tools††ccs:Information systems Database management systems††ccs:Computer systems organization Real\-time systems## 1\.Introduction: The Metrology Failure

### 1\.1\.The Wrong Unit

Energy\-per\-inference has become the de facto unit of AI energy accountability\(Strubellet al\.,[2019](https://arxiv.org/html/2605.22883#bib.bib1); Patterson and others,[2021](https://arxiv.org/html/2605.22883#bib.bib2); Oviedoet al\.,[2025](https://arxiv.org/html/2605.22883#bib.bib23); Jeghamet al\.,[2025](https://arxiv.org/html/2605.22883#bib.bib24); Chunget al\.,[2025](https://arxiv.org/html/2605.22883#bib.bib62)\)\. For classical batch workloads, one prompt in and one response out, the unit is coherent because inference count is fixed by the task definition\.

In agentic systems, this assumption breaks: a single user goal may trigger multi\-step orchestration, conditional tool calls, retry sequences, and failure\-recovery cycles\(Oviedoet al\.,[2025](https://arxiv.org/html/2605.22883#bib.bib23); Chienet al\.,[2023](https://arxiv.org/html/2605.22883#bib.bib25)\)whose depth is determined at runtime rather than by specification\. As a result, inference count reflects implementation behavior rather than task structure\.

Normalizing by inference count measures the cost of individual execution steps, not the cost of completing a goal\. This creates a mismatch between measurement unit and task semantics\. When inference count is used to normalize energy in systems where inference count itself is an implementation variable, the unit fails this requirement\. We refer to this as a unit misalignment\.

### 1\.2\.The Forcing Example

Consider two systems solving the same task\. System A succeeds on its first inference attempt\. System B fails four attempts before succeeding on the fifth \(Figure[1](https://arxiv.org/html/2605.22883#S1.F1)\)\. The resulting5×5\\timesamplification is a controlled thought experiment\.111The forcing example is a controlled\-thought experiment intended to expose the structural failure of inference\-level accounting under retry\. Section[8](https://arxiv.org/html/2605.22883#S8)reports empirically observed amplification on real workflows; the headline result \(4\.33×4\.33\\times, see §[8](https://arxiv.org/html/2605.22883#S8)\) is the quantity that should be cited by readers seeking a real measurement\.For clarity of exposition, assume each inference step has comparable energy costEinfE\_\{\\mathrm\{inf\}\}under identical hardware conditions\. In practice, failed attempts consume more energy than successful ones because the model exhausts more computation before producing an invalid output, making the real divergence larger than the5×5\\timesshown here\.

Under energy\-per\-inference accounting, both systems appear equivalent, since only successful or executed inference steps are counted uniformly\. However, total workflow energy differs: System A consumesEinfE\_\{\\mathrm\{inf\}\}, while System B consumes approximately5​Einf5E\_\{\\mathrm\{inf\}\}\.

This discrepancy arises because inference\-level accounting does not attribute failed attempts or retry sequences to the originating user goal\. In contrast,EpG\(Energy per Successful Goal\) aggregates energy across all attempts associated with goal completion, capturing this divergence explicitly\.OOI\(Orchestration Overhead Index\) expresses this gap as a ratio relative to a matched linear baseline; both are formally defined in Sections[6](https://arxiv.org/html/2605.22883#S6)and[7](https://arxiv.org/html/2605.22883#S7)\.

AEinfE\_\{\\inf\}✓E/inf:EinfE\_\{\\inf\}EpG:EinfE\_\{\\inf\}BEinfE\_\{\\inf\}×\\timesEinfE\_\{\\inf\}×\\timesEinfE\_\{\\inf\}×\\timesEinfE\_\{\\inf\}×\\timesEinfE\_\{\\inf\}✓E/inf:EinfE\_\{\\inf\}EpG:5​Einf5E\_\{\\inf\}same5×5\\timesEpGdivergenceFigure 1\.Two systems with identical energy\-per\-inference diverge by5×5\\timesunderEpG; see §[8](https://arxiv.org/html/2605.22883#S8)for measured aggregates\. Failed attempts are invisible to inference\-level accounting;EpGcaptures them by construction\.
### 1\.3\.A Real Agentic Trace

The forcing example is illustrative\. Consider a concrete agentic execution on a GSM8K arithmetic problem \(Figure[2](https://arxiv.org/html/2605.22883#S1.F2)\)\. The system issues a planning call, dispatches two tool invocations, encounters a JSON parse failure, retries, and synthesizes the answer\.

Goal: “Solve GSM8K arithmetic problem”\(exp\. 946, agentic, llama\_cpp, failure\_injection\)Workflow Unit: one episode, one retryAttempt 1controlled failure injected2,256 JwastedAttempt 2✓ success1,358 JusefulretryEpG=2,256\+1,358=𝟑,𝟔𝟏𝟒​𝐉\\textsc\{EpG\}=2\{,\}256\+1\{,\}358=\\mathbf\{3\{,\}614\\,J\}\(1 successful goal\)62\.4% ofEpGwastedon failed attemptE/inf counts only this:1,358 J – misses 62%Figure 2\.Goal, workflow unit, and retry accumulation on a real agentic run \(exp\. 946, GSM8K\-B, llama\_cpp, failure\_injection\)\. One goal maps to one workflow unit with one retry: a failed attempt \(2,256 J\) followed by a successful attempt \(1,358 J\)\.EpG= 3,614 J accumulates both; energy\-per\-inference counts only the successful attempt, missing 62\.4% of true cost\.The retry phase consumes more energy than the successful attempt that follows it: in the canonical trace, the failed attempt draws 2256\.1 J against 1358\.4 J for the successful attempt, yet inference\-level benchmarks exclude failed attempts entirely by construction\. During the LLM API wait phase in remote\-inference runs, overall task power drops to1\.01\.0W compared with0\.20\.2W during active planning\(Table[7](https://arxiv.org/html/2605.22883#S8.T7)\), a reduction that remains invisible without phase\-level measurement\(Jeghamet al\.,[2025](https://arxiv.org/html/2605.22883#bib.bib24); Kneese and Young,[2024](https://arxiv.org/html/2605.22883#bib.bib26)\)\. Methods that allocate energy purely by elapsed duration therefore overstate inference cost while understating orchestration and coordination overhead\. These distortions compound across multi\-step agentic workflows\.

### 1\.4\.Four Problems with Current Measurement

Energy\-per\-inference is not a valid unit for multi\-step agentic workflows\. A planner issuing four LLM calls and two retries is not “six inferences worth of energy”: it is one goal\-level computation whose energy is distributed across planning, execution, waiting, and recovery phases\. Four distinct failure modes arise in current practice\.

The unit problem\.Energy\-per\-inference counts implementation steps, not goal completions\. A system that retries four times before succeeding consumes5×5\\timesthe energy of one that succeeds immediately, yet both report identical energy\-per\-inference\.

The boundary problem\.Tools that allocate energy as TDP×\\timeswall\-time include post\-task framework teardown in the measurement window\. Teardown is a fixed absolute cost shared by both workflow types\. Because linear workflows complete faster, this fixed cost represents a larger fraction of reported linear energy than of reported agentic energy — inflating the denominator ofOOImore than the numerator, driving the ratio toward1\.0×1\.0\\timesregardless of the true orchestration overhead\(Schmidt and others,[2022](https://arxiv.org/html/2605.22883#bib.bib5); Google Cloud,[2024](https://arxiv.org/html/2605.22883#bib.bib44)\)\.

The attribution problem\.Raw package energy conflates idle system draw, concurrent process activity, and workload\-induced consumption\. Without explicit baseline subtraction and CPU\-fraction isolation, the measurement reflects the machine, not the task\(Thamm,[2025](https://arxiv.org/html/2605.22883#bib.bib27)\)\.

The reproducibility problem\.Without binding measurements to hardware identity, governor policy, and runtime configuration, the same workload on the same machine produces different numbers across runs, making cross\-paper comparison meaningless\(Mytkowiczet al\.,[2009](https://arxiv.org/html/2605.22883#bib.bib47); Pineauet al\.,[2021](https://arxiv.org/html/2605.22883#bib.bib48)\)\.

Sections[3](https://arxiv.org/html/2605.22883#S3)through[5](https://arxiv.org/html/2605.22883#S5)address each in turn; Section[6](https://arxiv.org/html/2605.22883#S6)defines the goal\-level unit that resolves the unit misalignment\.

### 1\.5\.Why Now

The measurement failures described above are not abstract concerns, and they are not future risks\. No published agentic AI benchmark, to our knowledge, normalizes energy by successfully completed goals rather than inference invocations\(Oviedoet al\.,[2025](https://arxiv.org/html/2605.22883#bib.bib23); Jeghamet al\.,[2025](https://arxiv.org/html/2605.22883#bib.bib24)\)\. The mismatch is fundamental: inference\-level normalization conflates implementation steps with goal completions, hiding retry amplification, orchestration overhead, and failure recovery costs\. Agentic AI deployments are scaling at a pace that makes the absence of a valid energy unit increasingly problematic\. Global AI inference demand is projected to grow from 15 TWh in 2025 to 347 TWh by 2030\(Schneider Electric,[2025](https://arxiv.org/html/2605.22883#bib.bib52)\), and agentic systems consume substantially more energy per interaction than single\-turn inference workloads\(Chen and others,[2026](https://arxiv.org/html/2605.22883#bib.bib17); Allianz Research,[2026](https://arxiv.org/html/2605.22883#bib.bib53)\)\. The shift from single\-pass inference to closed\-loop iterative reasoning changes the primary energy bottleneck from computation to orchestration overhead\(Chen and others,[2026](https://arxiv.org/html/2605.22883#bib.bib17)\), a change that inference\-level benchmarks are structurally blind to\.

Regulatory timelines compound the urgency\. The EU AI Act entered into force in August 2024 and reaches full enforcement in August 2026, with energy reporting obligations for General Purpose AI models and high\-risk AI systems\(European Parliament and Council of the European Union,[2024](https://arxiv.org/html/2605.22883#bib.bib54); White & Case LLP,[2025](https://arxiv.org/html/2605.22883#bib.bib55)\)\. The EU data\-centre energy efficiency package, expected in early 2026, will mandate per\-system energy KPIs\(European Parliamentary Research Service,[2025](https://arxiv.org/html/2605.22883#bib.bib56)\)\. California’s SB 253 requires Scope 1–3 emissions disclosure for companies over $1 billion in revenue starting 2026\(Schneider Electric,[2025](https://arxiv.org/html/2605.22883#bib.bib52)\)\. These frameworks require energy metrics that are reproducible, comparable across systems, and aligned with user\-visible outcomes\. Energy\-per\-inference satisfies none of these requirements for agentic workloads\.

This issue also appears in deployed systems\. A 280\-fold decline in inference costs since 2022 has driven a rebound effect: cheaper inference enables more agentic deployments, each consuming more energy per user goal than the single\-shot queries they replace\. Efficiency gains at the inference level cannot offset this growth\(Allianz Research,[2026](https://arxiv.org/html/2605.22883#bib.bib53)\)\. Without a goal\-level energy unit, neither system designers nor regulators can measure the true cost of agentic AI at the scale it is now operating\.

### 1\.6\.The PUE Analogy

The datacenter community faced a similar unit misalignment in the early era of large\-scale computing\. Power Usage Effectiveness \(PUE\) emerged as a practical remedy: operationally defined, empirically measurable, and widely adopted despite well\-known opportunities for metric gaming\(Barroso and Hölzle,[2007](https://arxiv.org/html/2605.22883#bib.bib13); The Green Grid,[2012](https://arxiv.org/html/2605.22883#bib.bib14)\)\.

EpGandOOIplay complementary roles for agentic AI\.EpGis an absolute unit, joules per successful goal, reportable for any system in isolation\.OOIis a comparative ratio requiring a matched linear baseline; EpG and OOI play analogous roles to PUE, as both isolate a specific overhead layer relative to a productive baseline, facility overhead in PUE and orchestration overhead inOOI\. Alternative datacenter metrics such as WUE normalize a different resource against energy output and do not share this ratio structure\. Like PUE,OOIis most meaningful when measurement conditions are controlled and reported\(Jeghamet al\.,[2025](https://arxiv.org/html/2605.22883#bib.bib24); Chienet al\.,[2023](https://arxiv.org/html/2605.22883#bib.bib25)\)\.

### 1\.7\.Contributions

1. \(1\)EpG\(Energy per Successful Goal\)\(§[6](https://arxiv.org/html/2605.22883#S6)\): A goal\-level energy unit that normalizes total workflow energy, including failed attempts and retries, by the number of successfully completed goals\.
2. \(2\)Measurement procedure\(§[3](https://arxiv.org/html/2605.22883#S3), §[4](https://arxiv.org/html/2605.22883#S4)\): a temporal boundary model \(t0/t1/t2t\_\{0\}/t\_\{1\}/t\_\{2\}\) combined with a five\-layer attribution hierarchy \(L0–L4\) that mapsRAPLsignals to workflow\-level energy, tagging each quantity asmeasured,calculated, orinferredat every layer\.
3. \(3\)Three\-hash reproducibility protocol\(§[5](https://arxiv.org/html/2605.22883#S5)\):ℋhw\\mathcal\{H\}\_\{\\mathrm\{hw\}\}\(hardware fingerprint\),ℋenv\\mathcal\{H\}\_\{\\mathrm\{env\}\}\(software environment\), andℋrun\\mathcal\{H\}\_\{\\mathrm\{run\}\}\(execution state\) jointly bind each measurement to its exact hardware and software context, enabling reproducible energy reporting\.
4. \(4\)OOI\(Orchestration Overhead Index\)\(§[7](https://arxiv.org/html/2605.22883#S7)\): a normalized ratio comparing agentic and linear workflows under identical task definitions and success criteria, isolating the energy cost of orchestration itself\.
5. \(5\)Empirical validation\(§[8](https://arxiv.org/html/2605.22883#S8)\): systematic evaluation across five claims — measurement validity, reproducibility, boundary correctness, discriminative power, and orchestration dominance — demonstratingOOI=4\.33×=4\.33\\timesacross827827\{\}agentic goals, with tool\-task inversion confirming directional correctness of the metric\.

## 2\.Measurement Ontology

Before defining how to measure agentic energy, we define*what*is being measured\. Three primitives constitute the minimum ontological commitments required to support a goal\-level energy unit\. They are not engineering choices specific toA\-LEMS; any apparatus claiming to measure goal\-level energy must adopt definitions of this kind\.

### 2\.1\.Goal

A*goal*is one unit of user intent, corresponding to one prompt\-answer pair as the user perceives it\. The goal is defined by what the user asks for and what the evaluation function accepts as a correct answer; it is independent of how the system internally satisfies it\. The number of inferences, tool calls, retries, or intermediate planning steps the system performs is implementation behavior, not part of the goal\.EpGis defined at the atomic level: a single user intent evaluated by a well\-defined success predicate under a fixed measurement boundary\.

A single user prompt is one goal regardless of its internal structure\. The arithmetic prompt “Solve: 23 \+ 47” is one goal\. The composite prompt “Plan a 3\-day Tokyo itinerary, then output it as JSON” is also one goal: the system either produces the requested JSON itinerary that the evaluation function accepts \(success\) or it does not \(failure\)\. Sub\-questions within a single prompt do not split into separate goals; they are internal task structure that the system must address to deliver the single accepted answer\.

Two cases require explicit treatment\. First, when a benchmark asks forKKindependent answers viaKKseparate prompts \(e\.g\., a 100\-question GSM8K subset\), those areKKgoals\. Second, when a single prompt contains multiple semantically independent questions and the evaluation function requires all to be answered \(e\.g\., “answer the following three questions: a, b, c”\), this is one goal whose evaluation is conjunctive: success requires all sub\-answers to be correct\.

Concretely, the goal granularity in this paper is determined by the benchmark specification \(one row of the GSM8K dataset = one goal\), not by the system’s internal decomposition\. Table[5](https://arxiv.org/html/2605.22883#S8.T5)lists all task families used in this paper and their corresponding goal definitions\.

### 2\.2\.Workflow Unit

A goal as defined in Section[2\.1](https://arxiv.org/html/2605.22883#S2.SS1)is a specification: it says what the user wants\. A*workflow unit*is the corresponding measured artifact: the actual sequence of system activity executed in service of one goal, with its energy, timing, and outcome recorded\(Figure[2](https://arxiv.org/html/2605.22883#S1.F2)\)\.

The relationship is one\-to\-one in the simplest case\. When the system succeeds on the first attempt with no retries, the workflow unit contains exactly one attempt, and the workflow unit aligns naturally with the goal: one prompt, one execution, one outcome\. In this case the distinction is mostly bookkeeping\.

Retry behavior makes this distinction important\. When a system retries after a JSON parse error, tool failure, hallucination filter, or timeout, the goal remains unchanged but the workflow unit contains multiple attempts: one or more failed attempts followed by either success or abandonment\. Each attempt consumes energy, so workflow energy is defined as the sum across all attempts:

Eworkflow=∑i=1nEattempt,iE\_\{\\mathrm\{workflow\}\}=\\sum\_\{i=1\}^\{n\}E\_\{\\mathrm\{attempt\},i\}
wherennis the number of attempts in this workflow unit \(n≥1n\\geq 1\)\. The goal\-success indicator𝟙success\\mathds\{1\}\_\{\\text\{success\}\}is11if the final attempt succeeded and0otherwise; Every workflow unit is counted regardless of outcome; only successful ones contribute to the successful\-goal count\. Figure[2](https://arxiv.org/html/2605.22883#S1.F2)illustrates this accumulation on a real agentic run \(exp\. 946, GSM8K\-B\): one failed attempt wastes 2,256 J before a successful retry adds 1,358 J, givingEpG=3,614=3\{,\}614J for one completed goal\.

This behavior arises becauseEpGcaptures retry\-driven energy amplification: the workflow unit accumulates attempts in the numerator while the denominator counts only the goals that ultimately succeeded\. A workflow unit that succeeds on its fifth attempt contributes five attempts’ worth of energy and one successful goal; a workflow unit that succeeds on its first attempt contributes one attempt’s worth of energy and one successful goal\. The5×5\\timesratio between them is invisible to inference\-level accounting and exactly visible toEpG\.

### 2\.3\.Successful Completion

Given the definitions of goal and workflow unit, the remaining ontological commitment is when a delivered result counts as*success*\.

A goal is*successfully completed*when the system’s final output is accepted by an evaluation function specified in advance of execution\. The pre\-specification requirement is critical: post\-hoc redefinition of what counts as success would allow systems to gameEpGby retroactively widening the success criterion until any delivered output qualifies\.

The evaluation function depends on the benchmark\. For GSM8K\(Cobbe and others,[2021](https://arxiv.org/html/2605.22883#bib.bib12)\), the function is exact\-match comparison against the ground\-truth integer answer extracted from the dataset\. For tool\-graph tasks, the function is a deterministic validator that checks whether the resulting state matches the specification\. For science\-QA tasks with multiple acceptable phrasings, the function is a normalized string match against any reference answer in the accepted set\. Table[5](https://arxiv.org/html/2605.22883#S8.T5)lists all task families used in this paper alongside their evaluation functions and success criteria\.

Three edge cases are resolved by convention: \(a\) success is binary — the evaluation function returns0or11and there is no partial credit at the goal level \(b\) if a workflow unit produces multiple candidate outputs \(e\.g\., self\-consistency sampling\), success requires only that at least one output is accepted, but all outputs’ energy is counted in the workflow unit; \(c\) a workflow unit that times out, exceeds the retry budget, or otherwise terminates without an accepted output counts as*failed*, not omitted: its energy is counted in the numerator and contributes zero to the successful\-goal denominator\. This ensures that giving up to save energy does not improve the reportedEpG\.

## 3\.Measurement Boundary Model

Energy is an integral over time,only well\-defined once integration limits are fixed\.*Boundary choice is a metrological commitment\.*Inconsistent boundaries cause systematic error that compounds across replications\.

### 3\.1\.Execution Boundaries

Every energy measurement implicitly commits to an integration window; without an explicit boundary, reported values are incomparable across tools and runs\.A\-LEMSdefines three ordered boundaries that partition each run into distinct intervals \(Figure[3](https://arxiv.org/html/2605.22883#S3.F3)\):

tpret\_\{\\mathrm\{pre\}\}t0t\_\{0\}t1t\_\{1\}t2t\_\{2\}Pre\-taskdiagnosticAttribution window\[t0,t1\]\[t\_\{0\},t\_\{1\}\]Post\-taskexcludedsetupexecution \+ retriesRAPL counter read\(context baseline\)start\_measurement\(\)attribution beginsexecutor returnsattribution endsETL \+ DB writescompleteFigure 3\.Measurement boundary model\. Three ordered anchorstpret\_\{\\mathrm\{pre\}\},t0t\_\{0\},t1t\_\{1\},t2t\_\{2\}partition each run into pre\-task diagnostic, attributed task\[t0,t1\]\[t\_\{0\},t\_\{1\}\], and post\-task excluded windows\.”

Total run energy follows directly from two hardware counter reads att0t\_\{0\}andt1t\_\{1\}:Etask=RAPL​\(t1\)−RAPL​\(t0\)E\_\{\\mathrm\{task\}\}=\\mathrm\{RAPL\}\(t\_\{1\}\)\-\\mathrm\{RAPL\}\(t\_\{0\}\)\. This value is exact and requires no intermediate sampling\. An agentic workflow, however, spans structurally distinct phases: planning, LLM execution, tool dispatch, and synthesis\. Understanding which phase drives energy consumption is central to the thesis\. IntermediateRAPLsamples within\[t0,t1\]\[t\_\{0\},t\_\{1\}\]provide the temporal resolution needed to attributeEtaskE\_\{\\mathrm\{task\}\}across these phases\. Without phase\-level samples, the total energy is correct but opaque — we know how much was consumed but not by which phase\. Since orchestration dominance is the central empirical claim, phase attribution is essential\. Coverage𝒞\\mathcal\{C\}measures how completely the sample stream spans the attribution window, and therefore how reliably phase boundaries can be resolved\. Concretely, eachRAPLsample spans an interval\[si,ei\]\[s\_\{i\},e\_\{i\}\]within\[t0,t1\]\[t\_\{0\},t\_\{1\}\]; at 100 Hz the sampler collects approximately 103\.0 samples per second, each covering≈\\approx10 ms \(empirically validated in Section[8\.2](https://arxiv.org/html/2605.22883#S8.SS2)\)\. Coverage is the fraction of\[t0,t1\]\[t\_\{0\},t\_\{1\}\]these intervals collectively span:

\(1\)C=\|\(⋃i\[si,ei\]\)∩\[t0,t1\]\|t1−t0×100%C=\\frac\{\\left\|\\left\(\\bigcup\_\{i\}\[s\_\{i\},e\_\{i\}\]\\right\)\\cap\[t\_\{0\},t\_\{1\}\]\\right\|\}\{t\_\{1\}\-t\_\{0\}\}\\times 100\\,\\%A run with lowCCreports correct total energy but leaves phase boundaries unresolved — the energy is accounted for but not attributed\. At gold tier \(C≥95%C\\geq 95\\%\), at most 5% of the execution window lacks a sample, bounding the maximum unresolved phase interval to 332\.6 ms observed across all 2228 runs\. Runs belowC=80%C=80\\%are excluded from phase\-level analyses; the full empirical distribution over the canonical dataset is reported in Section[8\.2](https://arxiv.org/html/2605.22883#S8.SS2)\.

### 3\.2\.Workflow Energy

Each run decomposes into three non\-overlapping energy components:

\(2\)Eworkflow=∫t0t1p​\(t\)​𝑑t⏟Etask\+∫tpret0p​\(t\)​𝑑t⏟Epre\+∫t1t2p​\(t\)​𝑑t⏟EpostE\_\{\\mathrm\{workflow\}\}=\\underbrace\{\\int\_\{t\_\{0\}\}^\{t\_\{1\}\}p\(t\)\\,dt\}\_\{E\_\{\\mathrm\{task\}\}\}\+\\underbrace\{\\int\_\{t\_\{\\mathrm\{pre\}\}\}^\{t\_\{0\}\}p\(t\)\\,dt\}\_\{E\_\{\\mathrm\{pre\}\}\}\+\\underbrace\{\\int\_\{t\_\{1\}\}^\{t\_\{2\}\}p\(t\)\\,dt\}\_\{E\_\{\\mathrm\{post\}\}\}
EtaskE\_\{\\mathrm\{task\}\}is the primary attribution target and the sole component enteringEpG\(Section[6](https://arxiv.org/html/2605.22883#S6)\)\.EpreE\_\{\\mathrm\{pre\}\}captures system activity during the pre\-task diagnostic window;EpostE\_\{\\mathrm\{post\}\}covers framework teardown after executor return\. Both are recorded for diagnostic purposes and excluded fromEpG\. How each component is computed from hardware signals is defined in Section[4](https://arxiv.org/html/2605.22883#S4); empirical magnitudes are reported in Section[8\.4](https://arxiv.org/html/2605.22883#S8.SS4)\.

### 3\.3\.Boundary Failure Modes

The decomposition in Equation[2](https://arxiv.org/html/2605.22883#S3.E2)makes boundary choice concrete; tools that omit this commitment conflateEpreE\_\{\\mathrm\{pre\}\},EpostE\_\{\\mathrm\{post\}\}, or concurrent process energy withEtaskE\_\{\\mathrm\{task\}\}resulting in four failure modes explained in Table[1](https://arxiv.org/html/2605.22883#S3.T1)\.

Table 1\.Boundary failure modes, their mechanisms, andA\-LEMSmitigations\. All four modes introduce systematic bias independent of workload; the\[t0,t1\]\[t\_\{0\},t\_\{1\}\]attribution boundary eliminates all four\. OOI impact: direction of distortion on theOOIratio if unmitigated\.Truncation bias is inherent to single\-shot inference benchmarks\(Mattson and others,[2020](https://arxiv.org/html/2605.22883#bib.bib11)\), which predate agentic multi\-step execution\. The remaining three modes affect any tool that does not define an explicit\[t0,t1\]\[t\_\{0\},t\_\{1\}\]attribution window\(Schmidt and others,[2022](https://arxiv.org/html/2605.22883#bib.bib5); Thamm,[2025](https://arxiv.org/html/2605.22883#bib.bib27)\)\. Together they can produce order\-of\-magnitude errors on short workloads where idle draw dominates and framework overhead is non\-negligible relative to task energy\.

## 4\.A\-LEMS: Five\-Layer Energy Observation Model

Raw hardware energy counters are necessary but not sufficient for workflow\-level energy attribution\. A singleRAPLread spanning a complete agentic run conflates idle system draw, background process activity, and the target workflow into one undifferentiated number\. Separating these contributions requires a sequence of transformations, each addressing a distinct observability gap that the previous layer cannot resolve\.A\-LEMSformalizes this sequence as five layers, and illustrated with canonical trace values in Figure[4](https://arxiv.org/html/2605.22883#S4.F4)and Table[9](https://arxiv.org/html/2605.22883#A4.T9)\. The current implementation measures the local CPU package; the fullEpGdefinition in Section[6](https://arxiv.org/html/2605.22883#S6)encompasses GPU, NIC, and remote inference energy, and this paper reports the local\-CPU component\.

𝐋𝟎\\mathbf\{L\_\{0\}\}: HardwareEpkg=Ecore\+Euncore\+EdramE\_\{\\mathrm\{pkg\}\}=E\_\{\\mathrm\{core\}\}\+E\_\{\\mathrm\{uncore\}\}\+E\_\{\\mathrm\{dram\}\}measured𝐋𝟏\\mathbf\{L\_\{1\}\}: Baseline isolationEdyn=max⁡\(0,Epkg−Pbase⋅Δ​t\)E\_\{\\mathrm\{dyn\}\}=\\max\(0,\\;E\_\{\\mathrm\{pkg\}\}\-P\_\{\\mathrm\{base\}\}\\cdot\\Delta t\)calculated𝐋𝟐\\mathbf\{L\_\{2\}\}: Process attributionEattr=fcpu⋅EdynE\_\{\\mathrm\{attr\}\}=f\_\{\\mathrm\{cpu\}\}\\cdot E\_\{\\mathrm\{dyn\}\},fcpu=Δ​tickspid/Δ​tickstotalf\_\{\\mathrm\{cpu\}\}=\\Delta\\mathrm\{ticks\}\_\{\\mathrm\{pid\}\}/\\Delta\\mathrm\{ticks\}\_\{\\mathrm\{total\}\}calculated𝐋𝟑\\mathbf\{L\_\{3\}\}: Phase decompositionEattr=Eplan\+Eexec\+Esyn\+EgapE\_\{\\mathrm\{attr\}\}=E\_\{\\mathrm\{plan\}\}\+E\_\{\\mathrm\{exec\}\}\+E\_\{\\mathrm\{syn\}\}\+E\_\{\\mathrm\{gap\}\}calculated𝐋𝟒\\mathbf\{L\_\{4\}\}: Goal aggregationEpG=∑iEattr\(i\)/Nsuccess\\textsc\{EpG\}=\\sum\_\{i\}E\_\{\\mathrm\{attr\}\}^\{\(i\)\}/N\_\{\\mathrm\{success\}\}calculated4274\.1 J3827\.7 J3614\.5 J552\.6\+\+115\.2\+\+69\.2\+\+2877\.5 J3614\.5 JOOI=14\.2×=14\.2\\timesFigure 4\.Five\-layer attribution hierarchy with provenance tiers\. Right\-hand values trace a single canonical agentic run \(run 3343, exp 946,gsm8k\_basic,llama\_cpp\) through all five layers toEpG\. The gap term \(2877\.5 J\) captures retry and inter\-phase coordination energy, the dominant energy cost in agentic workflows\.### 4\.1\.L0 — Raw Hardware Energy

RAPL\(David and others,[2010](https://arxiv.org/html/2605.22883#bib.bib9); Intel Corporation,[2023](https://arxiv.org/html/2605.22883#bib.bib10)\)exposes package\-level energy via model\-specific registers \(MSRs\) at approximately 1 mJ resolution\(Thamm,[2025](https://arxiv.org/html/2605.22883#bib.bib27); Koszczałet al\.,[2025](https://arxiv.org/html/2605.22883#bib.bib28)\):

\(3\)Epkg=Ecore\+Euncore\+EdramE\_\{\\mathrm\{pkg\}\}=E\_\{\\mathrm\{core\}\}\+E\_\{\\mathrm\{uncore\}\}\+E\_\{\\mathrm\{dram\}\}
The counters are updated by the processor independently of user\-space activity and are not influenced by application\-level instrumentation or scheduling decisions\. This makes L0 the physical grounding signal from which all subsequent transformations derive\. The measurement scope is the local CPU package and excludes GPU execution, network interface energy, and remote inference\(Koszczałet al\.,[2025](https://arxiv.org/html/2605.22883#bib.bib28); Niazet al\.,[2025](https://arxiv.org/html/2605.22883#bib.bib29); Kulkarni and others,[2026](https://arxiv.org/html/2605.22883#bib.bib31)\)\. This scope is intentional and bounded: the fullEpGdefinition \(§[6](https://arxiv.org/html/2605.22883#S6)\) encompasses\(ECPU\+EGPU\)local\+ENIC\+Eremote\(E\_\{\\mathrm\{CPU\}\}\+E\_\{\\mathrm\{GPU\}\}\)\_\{\\mathrm\{local\}\}\+E\_\{\\mathrm\{NIC\}\}\+E\_\{\\mathrm\{remote\}\}; this paper instantiates and validates the local\-CPU component, which is the dominant term for CPU\-orchestrated agentic workloads running local inference\. This signal ismeasureddirectly from hardware counters with no software transformation\.

### 4\.2\.L1 — Baseline\-Subtracted Energy

Computing systems draw non\-negligible power even at idle —2\.262\.26W on our platform — a well\-documented non\-proportionality\(Barroso and Hölzle,[2007](https://arxiv.org/html/2605.22883#bib.bib13); Aquino\-Brítezet al\.,[2025](https://arxiv.org/html/2605.22883#bib.bib30); Bertschet al\.,[2025](https://arxiv.org/html/2605.22883#bib.bib32)\)\. Without correcting for this floor, short linear runs accumulate proportionally more idle contamination than longer agentic runs, compressing OOI toward 1 and understating orchestration overhead\. The correction is:

\(4\)Edyn=max⁡\(0,Epkg−Pbaseline⋅Δ​t\)E\_\{\\mathrm\{dyn\}\}=\\max\\\!\\bigl\(0,\\;E\_\{\\mathrm\{pkg\}\}\-P\_\{\\mathrm\{baseline\}\}\\cdot\\Delta t\\bigr\)
wherePbaselineP\_\{\\mathrm\{baseline\}\}is the mean idle power estimated over multiple non\-overlapping sampling windows, with samples beyond two standard deviations of the mean excluded before averaging\. Baseline sampling runs under CPU affinity constraints to prevent scheduler\-induced core migration from inflating the idle estimate\. On our platform, this correction removes approximately 120 J from a typical agentic run\. Provenance:calculatedfrommeasuredinputs via closed\-form arithmetic\.

### 4\.3\.L2 — Process\-Level Energy Attribution

\(5\)Eattr=fcpu⋅Edyn,fcpu=Δ​tickspidΔ​tickstotalE\_\{\\mathrm\{attr\}\}=f\_\{\\mathrm\{cpu\}\}\\cdot E\_\{\\mathrm\{dyn\}\},\\quad f\_\{\\mathrm\{cpu\}\}=\\frac\{\\Delta\\mathrm\{ticks\}\_\{\\mathrm\{pid\}\}\}\{\\Delta\\mathrm\{ticks\}\_\{\\mathrm\{total\}\}\}
wherefcpuf\_\{\\mathrm\{cpu\}\}is the fraction of CPU time consumed by the target process, derived from/proc/\{pid\}/statand/proc/statcounter deltas over\[t0,t1\]\[t\_\{0\},t\_\{1\}\], representing the fraction of CPU time consumed by the target process\. Without this step, concurrent background processes contaminate the energy estimate; on a lightly loaded research machine the effect is modest, but it is not negligible and grows with system utilization\. The projection is a heuristic: CPU time is a widely accepted but imperfect proxy for energy consumption, and memory\- or I/O\-bound phases may carry energy not fully captured byfcpuf\_\{\\mathrm\{cpu\}\}\. The limitation is explicit and bounded by the invariantEattr≤EdynE\_\{\\mathrm\{attr\}\}\\leq E\_\{\\mathrm\{dyn\}\}, enforced at runtime\. Figure[5](https://arxiv.org/html/2605.22883#S4.F5)\(b\) confirms this invariant holds across all 827 canonical paired runs\. The projection iscalculatedvia a heuristic transform ofmeasuredCPU ticks\.

### 4\.4\.L3 — Phase\-Resolved Energy Decomposition

L2 produces a single attributed energy total per run\. That total is correct and sufficient for computingEpGandOOI, but it is opaque: planning, execution, synthesis, retry energy, and framework overhead all collapse into one number\. L3 partitionsEattrE\_\{\\mathrm\{attr\}\}across semantic phase windows using direct integration of hardware energy samples:

\(6\)Ephase=∑i∈𝒮ϕΔ​EiE\_\{\\mathrm\{phase\}\}=\\sum\_\{i\\in\\mathcal\{S\}\_\{\\phi\}\}\\Delta E\_\{i\}
where𝒮ϕ\\mathcal\{S\}\_\{\\phi\}denotes the set of sampled energy intervals whose timestamps overlap phase windowϕ\\phi, andΔ​Ei=Eend\(i\)−Estart\(i\)\\Delta E\_\{i\}=E^\{\(i\)\}\_\{\\mathrm\{end\}\}\-E^\{\(i\)\}\_\{\\mathrm\{start\}\}is the per\-interval RAPL delta\. Samples are collected at approximately 100 Hz; phase boundaries from workflow event logs are aligned with these intervals to recover phase\-specific energy\. The residual gap term absorbs all energy outside named phase windows and has two components: retry energy from failed attempts, and inter\-phase orchestration overhead between phases\. In the canonical trace run \(run 3343, exp 946\), the gap accounts for 2877\.5 J, of which 2256\.1 J is retry energy from the failed attempt 1 and 621\.4 J is inter\-phase coordination energy\. This decomposition is only possible because L3 measures phases independently from RAPL samples rather than allocating energy by time fraction\.

The alternative time\-fraction approach assigns energy proportional to phase duration, producing a systematic and measurable error\. On run 3343, time\-fraction would attribute planning 1387\.6 J against the measured 552\.6 J, execution 307\.6 J against 115\.2 J, synthesis 105\.5 J against 69\.2 J, and gap 1813\.7 J against 2877\.5 J\. The error is not random: synthesis draws 25\.9 W during its 2\.7 s window — below the run average of 39\.5 W — so time\-fraction over\-charges it, while the gap period draws 62\.7 W during retry execution, above the run average, causing time\-fraction to under\-charge it\. On this run, time\-fraction misattributes energy in both directions — overcharging synthesis and undercharging the gap — confirming that phase power is not proportional to phase duration: time\-fraction is not a valid method for phase\-level attribution\. Had time\-fraction been used, the gap’s share of attributed energy would appear as 50\.2% rather than the measured 79\.6%, understating the true cost of failed attempts by a measurable margin\. Table[2](https://arxiv.org/html/2605.22883#S4.T2)shows the full v1 vs v2 comparison for all four phases\.

Figure[5](https://arxiv.org/html/2605.22883#S4.F5)\(a\) shows the phase breakdown aggregated across all 827 canonical agentic goals: planning, synthesis, and gap together dominate, while active execution contributes the smallest share\. Figure[5](https://arxiv.org/html/2605.22883#S4.F5)\(b\) confirms the L1→\\toL2 monotonicity invariantEattr≤EdynE\_\{\\mathrm\{attr\}\}\\leq E\_\{\\mathrm\{dyn\}\}holds across all 827 paired runs\. Figure[5](https://arxiv.org/html/2605.22883#S4.F5)\(c\) shows the power profile of run 3343 with the v1 time\-fraction counterfactual overlaid: synthesis draws 25\.9 W during its 2\.7 s window while the gap period draws higher power due to retry execution, yet time\-fraction assigns both the same uniform 39\.5 W\. The integration iscalculateddirectly frommeasuredRAPL samples and event boundaries, with provenance recorded in the methodology registry \(§[4\.6](https://arxiv.org/html/2605.22883#S4.SS6)\)\.

![Refer to caption](https://arxiv.org/html/2605.22883v1/x1.png)Figure 5\.Mean energy per run decomposed by phase \(planning, execution, synthesis, gap\) across all 827 canonical agentic goals\. The gap term dominates, confirming that orchestration wait rather than active inference drives agentic energy consumption\.
### 4\.5\.L4 — Goal\-Level Aggregation

The four layers above operate at the run level\. L4 lifts the result to the goal level by summing attributed energy across all attempts belonging to a goal and dividing by the number of successfully completed goals:

\(7\)EpG=∑j=1AEattr\(j\)Nsuccess\\textsc\{EpG\}=\\frac\{\\sum\_\{j=1\}^\{A\}E\_\{\\mathrm\{attr\}\}^\{\(j\)\}\}\{N\_\{\\mathrm\{success\}\}\}
whereAAis the total number of attempts, including failures\. This is the step that makes retry energy visible\. A workflow that fails once before succeeding charges both attempts to the goal; a perfectly reliable workflow charges only one\. The denominatorNsuccessN\_\{\\mathrm\{success\}\}ties the metric to task completion rather than task invocation, aligning measurement with system utility\. The aggregation is acalculatedclosed\-form sum overmeasuredper\-attempt energy values\.

### 4\.6\.Signal Provenance and Measurement Tiers

Each layer is assigned a provenance tier drawn from theA\-LEMSmethodology registry:measuredfor quantities read directly from hardware counters,calculatedfor quantities derived from measured inputs through closed\-form or widely\-validated transforms, andinferredfor quantities derived through alignment of independently measured signals where the correspondence is approximate\. All five layers in this paper are eithermeasuredorcalculated; the tier assignments are stored machine\-readably in the methodology registry alongside each formula and are queryable without source code access\.

### 4\.7\.Composition Across Layers

The five layers are not alternatives; they execute in sequence for every measured run, with each layer’s output feeding directly into the next\. Table[2](https://arxiv.org/html/2605.22883#S4.T2)traces this computation for a single canonical run, showing how 3614\.5 J emerges from a raw hardware read of 4274\.1 J through four successive transformations\. Table[3](https://arxiv.org/html/2605.22883#S4.T3)documents the failure mode at each layer and its expected effect on reported metrics\.

Table 2\.Energy flow through all five layers for a single canonical agentic run \(run 3343, exp 946, task:gsm8k\_basic, provider:llama\_cpp, model:tinyllama\-1b\-gguf\)\. Attempt 1 failed under the failure\-injection protocol, consuming 2256\.1 J before attempt 2 succeeded\. Task simplicity is deliberate: it isolates orchestration overhead from task complexity\. v1 column shows time\-fraction counterfactual\. All values fromexperiments\.dbmacros\.LayerTierOperationv1 \(J\)v2 \(J\)L0meas\.Raw RAPLEpkgE\_\{\\mathrm\{pkg\}\}—4274\.1L1calc\.−\-idle baseline \(446\.4 J\)→Edyn\\rightarrow E\_\{\\mathrm\{dyn\}\}—3827\.7L2calc\.×fcpu→Eattr\\times\\,f\_\{\\mathrm\{cpu\}\}\\rightarrow E\_\{\\mathrm\{attr\}\}—3614\.5L3calc\.Planning \(35\.1 s\)1387\.6552\.6Execution \(7\.8 s\)307\.6115\.2Synthesis \(2\.7 s, 25\.9 W\)105\.569\.2Gap \(45\.9 s\): retry 2256\.1 J \+ coordination 621\.4 J1813\.72877\.5L4calc\.Attempt 1 failed \(2256\.1 J wasted\)Attempt 2 success \(1358\.4 J\)EpG= total / 1 goal3614\.5Linear baseline run 3342254\.5OOI \(this goal instance\)14\.2×\\timesTable 3\.Representative failure modes across the five\-layer observation stack and their effect on reported energy metrics\.

## 5\.Reproducibility Protocol

Energy measurements are sensitive to hardware state, firmware revision, kernel behavior, scheduler policy, thermal history, and software drift\(Mytkowiczet al\.,[2009](https://arxiv.org/html/2605.22883#bib.bib47); Thamm,[2025](https://arxiv.org/html/2605.22883#bib.bib27)\)\. A measurement that cannot be traced to its exact execution context cannot be reproduced and cannot be falsified\.A\-LEMSaddresses this through a three\-hash provenance protocol that binds every run to an immutable record of its hardware identity, software environment, and runtime measurement state\(Figure[6](https://arxiv.org/html/2605.22883#S5.F6)\)\.

hardware\_configcpu, microcode,kernel, RAPLebe69422environment\_configpython, OS,git commitMeasurement stategovernor, turbo,baseline\_idℋhw\\mathcal\{H\}\_\{\\mathrm\{hw\}\}ℋenv\\mathcal\{H\}\_\{\\mathrm\{env\}\}ℋrun\\mathcal\{H\}\_\{\\mathrm\{run\}\}runstable: all three hashes stored per measurement\+git\_dirtyflag exposed \(not hidden\)Figure 6\.Three\-hash reproducibility protocol\.ℋhw\\mathcal\{H\}\_\{\\mathrm\{hw\}\}encodes hardware fingerprint;ℋenv\\mathcal\{H\}\_\{\\mathrm\{env\}\}encodes software environment including git dirty flag;ℋrun\\mathcal\{H\}\_\{\\mathrm\{run\}\}encodes measurement state and includesℋhw\\mathcal\{H\}\_\{\\mathrm\{hw\}\}\. All three stored per run in therunstable\.\(8\)ℋhw\\displaystyle\\mathcal\{H\}\_\{\\mathrm\{hw\}\}=SHA256​\(Mcpu​‖Vμ‖​K∥DRAPL\)\\displaystyle=\\mathrm\{SHA256\}\(M\_\{\\mathrm\{cpu\}\}\\\|V\_\{\\mu\}\\\|K\\\|D\_\{\\mathrm\{RAPL\}\}\)\(9\)ℋenv\\displaystyle\\mathcal\{H\}\_\{\\mathrm\{env\}\}=SHA256​\(P​‖O‖​Gcommit​‖Gdirty‖​Fver∥Sschema\)\\displaystyle=\\mathrm\{SHA256\}\(P\\\|O\\\|G\_\{\\mathrm\{commit\}\}\\\|G\_\{\\mathrm\{dirty\}\}\\\|F\_\{\\mathrm\{ver\}\}\\\|S\_\{\\mathrm\{schema\}\}\)\(10\)ℋrun\\displaystyle\\mathcal\{H\}\_\{\\mathrm\{run\}\}=SHA256​\(Ggov​‖Tturbo‖​ℋhw​‖ℋenv‖​Bid\)\\displaystyle=\\mathrm\{SHA256\}\(G\_\{\\mathrm\{gov\}\}\\\|T\_\{\\mathrm\{turbo\}\}\\\|\\mathcal\{H\}\_\{\\mathrm\{hw\}\}\\\|\\mathcal\{H\}\_\{\\mathrm\{env\}\}\\\|B\_\{\\mathrm\{id\}\}\)whereMcpuM\_\{\\mathrm\{cpu\}\}is the CPU model string,VμV\_\{\\mu\}the microcode version,KKthe kernel release,DRAPLD\_\{\\mathrm\{RAPL\}\}the available RAPL domains,PPthe Python version,OOthe OS name,GcommitG\_\{\\mathrm\{commit\}\}the git commit hash,GdirtyG\_\{\\mathrm\{dirty\}\}the repository dirty flag,FverF\_\{\\mathrm\{ver\}\}the framework version,SschemaS\_\{\\mathrm\{schema\}\}the database schema version,GgovG\_\{\\mathrm\{gov\}\}the CPU frequency governor,TturboT\_\{\\mathrm\{turbo\}\}the turbo state, andBidB\_\{\\mathrm\{id\}\}the baseline measurement identifier; queryable fields and canonical values are in Table[4](https://arxiv.org/html/2605.22883#S5.T4)\.

The three hashes compose into a diagnostic ladder rather than a binary pass/fail gate\.ℋhw\\mathcal\{H\}\_\{\\mathrm\{hw\}\}fingerprints the physical substrate: CPU model, microcode revision, kernel version, and available RAPL domains\.ℋenv\\mathcal\{H\}\_\{\\mathrm\{env\}\}fingerprints the software stack, including git commit, dirty\-repository flag, and database schema version — ensuring attribution logic changes across migrations are captured in the environment fingerprint\.ℋrun\\mathcal\{H\}\_\{\\mathrm\{run\}\}incorporates bothℋhw\\mathcal\{H\}\_\{\\mathrm\{hw\}\}andℋenv\\mathcal\{H\}\_\{\\mathrm\{env\}\}transitively, adding governor state, turbo setting, and baseline identifier\. A mismatch onℋrun\\mathcal\{H\}\_\{\\mathrm\{run\}\}therefore has a precise interpretation: ifℋhw\\mathcal\{H\}\_\{\\mathrm\{hw\}\}matches andℋenv\\mathcal\{H\}\_\{\\mathrm\{env\}\}also matches, the hardware and software environment are identical and anyℋrun\\mathcal\{H\}\_\{\\mathrm\{run\}\}divergence is confined to runtime state fields \(governor,turbo,baseline\_id\), inspectable directly from therunstable\. Ifℋenv\\mathcal\{H\}\_\{\\mathrm\{env\}\}differs, the software environment changed — git commit, runtime versions, or database schema version — and the relevant fields are queryable fromenvironment\_config\. Ifℋhw\\mathcal\{H\}\_\{\\mathrm\{hw\}\}also differs, the physical substrate changed and the individually stored fields \(cpu\_model,rapl\_domains\) identify exactly what changed and by how much, serving as a structured compatibility checklist rather than a rejection signal\. All 827 agentic and 827 linear canonical runs shareℋhw=ebe694229b1b9d87\\mathcal\{H\}\_\{\\mathrm\{hw\}\}=\\texttt\{ebe694229b1b9d87\}, produced from a clean repository state \(Gdirty=0G\_\{\\mathrm\{dirty\}\}=0, commitc239acea674daefe\)\. Across 9 distinctℋenv\\mathcal\{H\}\_\{\\mathrm\{env\}\}values as git commits evolved between sessions, OOI remains immune: each agentic–linear pair executes within the same session under identicalℋenv\\mathcal\{H\}\_\{\\mathrm\{env\}\}andℋrun\\mathcal\{H\}\_\{\\mathrm\{run\}\}, so software drift divides out of the ratio by construction\. Table[4](https://arxiv.org/html/2605.22883#S5.T4)formalizes the diagnostic ladder; Table[6](https://arxiv.org/html/2605.22883#S8.T6)confirms all axes were controlled across the full 2228 runs\.

Table 4\.Reproducibility diagnostic semantics\. Canonical values are from theA\-LEMSdataset \(ℋhw=ebe694229b1b9d87\\mathcal\{H\}\_\{\\mathrm\{hw\}\}=\\texttt\{ebe694229b1b9d87\}, commitc239acea674daefe,Gdirty=0G\_\{\\mathrm\{dirty\}\}=0\)\.
## 6\.Energy per Successful Goal \(EpG\)

The preceding sections establish the measurement apparatus\. What remains is the unit itself\. LetEattempt,iE\_\{\\mathrm\{attempt\},i\}denote the energy consumed during attemptiiof a workflow, as obtained from the L2 process\-level projection \(Section[4](https://arxiv.org/html/2605.22883#S4)\)\. A workflow may span one or more attempts, each drawing real energy from the hardware regardless of whether it reaches a successful outcome\. The total energy attributable to workflowjjis the sum across all its attempts, successful or otherwise:

\(11\)Eworkflow\(j\)=∑i=1njEattempt,i\(j\)E\_\{\\mathrm\{workflow\}\}^\{\(j\)\}=\\sum\_\{i=1\}^\{n\_\{j\}\}E\_\{\\mathrm\{attempt\},i\}^\{\(j\)\}wherenjn\_\{j\}is the number of attempts in workflowjj\. Let𝒲\\mathcal\{W\}denote the set of all workflow units observed during an evaluation, and𝒲\+⊆𝒲\\mathcal\{W\}^\{\+\}\\subseteq\\mathcal\{W\}the subset of workflows that delivered a successful outcome\(Bertschet al\.,[2025](https://arxiv.org/html/2605.22883#bib.bib32); Afzalet al\.,[2023](https://arxiv.org/html/2605.22883#bib.bib33); Kulkarni and others,[2026](https://arxiv.org/html/2605.22883#bib.bib31)\)\. Energy per Successful Goal is then:

\(12\)EpG=∑j∈𝒲Eworkflow\(j\)\|𝒲\+\|\\boxed\{\\textsc\{EpG\}=\\frac\{\\sum\_\{j\\in\\mathcal\{W\}\}E\_\{\\mathrm\{workflow\}\}^\{\(j\)\}\}\{\|\\mathcal\{W\}^\{\+\}\|\}\}with unitjoules per successful goal \(J/goal\)\. This paper instantiatesEattemptE\_\{\\mathrm\{attempt\}\}as local CPU package energy measured viaRAPL; the temporal boundary model in Section[3](https://arxiv.org/html/2605.22883#S3)defines the attribution window with precision\. The definition is substrate\-agnostic: as measurement scope expands to GPU, network interface, and remote provider energy, the formula absorbs those terms without structural change\. Figure[2](https://arxiv.org/html/2605.22883#S1.F2)grounds this in a concrete agentic execution: two attempts summing to 3614\.5 J for one successful goal yieldEpG==3614\.5 J/goal, against a linear baseline of 254\.5 J for the same task, anOOIof 14\.2×\\timeson a single run\. The same metric that produces this amplification for reasoning tasks yieldsOOI<1\.0×<1\.0\\timesfor tool\-dispatch tasks, where agentic execution replaces costly token generation with a direct function call\. That directional sensitivity is a property of the definition, not a tuning parameter, and it is validated systematically in Section[8](https://arxiv.org/html/2605.22883#S8)\. The denominator behavior under representative workflow structures is shown in Figure[7](https://arxiv.org/html/2605.22883#S6.F7)\.

\(a\) Linearsingle attempt\(b\) Agenticwith retry\(c\) Aggregaten=827n=827\{\}goals254\.5 J✓ successEpG=254\.5​J1​goal\\textsc\{EpG\}=\\frac\{254\.5\\,\\mathrm\{J\}\}\{1\\;\\mathrm\{goal\}\}=254\.5=254\.5J/goal2256\.1 J×\\timesfailed1358\.4 J✓ successEpG=2256\.1\+1358\.4​J1​goal\\textsc\{EpG\}=\\frac\{2256\.1\+1358\.4\\,\\mathrm\{J\}\}\{1\\;\\mathrm\{goal\}\}=3614\.5=3614\.5J/goal OOI=14\.2×=14\.2\\timesfailed\>\>success205\.3 J888\.1 JLinearAgenticMean OOI=4\.33×=4\.33\\times\(n=827n=827\{\}\)Energy \(J\), proportional barsFigure 7\.EpGdenominator behavior with real measured energy values\. \(a\) Linear baseline: single successful attempt at 254\.5 J/goal\. \(b\) Agentic failure\-injected run \(exp\. 946, run 3343\): the failed attempt \(2256\.1 J\) and successful attempt \(1358\.4 J\) both enter the numerator; one goal enters the denominator, yieldingEpG==3614\.5 J/goal andOOI=14\.2×=14\.2\\times\. Inference\-level accounting assigns identical cost to both attempts\. \(c\) Aggregate across 827 canonical goals: meanEpGof 888\.1 J \(agentic\) vs\. 205\.3 J \(linear\),OOI=4\.33×=4\.33\\times\. Each panel uses an independent energy scale to preserve readability across the three\-order\-of\-magnitude range\.The structure of the formula encodes a deliberate metrological choice that is worth making explicit\. Failed attempts enter the numerator because they consumed real energy on behalf of the goal: the hardware ran, the model executed, and joules were drawn from the wall, regardless of the outcome\. Those attempts do not increment the denominator because no goal was delivered to the user\. A system that requires four attempts before succeeding reports anEpGfour times higher than one that succeeds immediately, even under identical per\-attempt energy cost\. This property makesEpGa behavioral signal as much as an energy signal: two systems with identical per\-attempt energy but different reliability profiles will separate cleanly underEpG, enabling researchers to study how model behavior — prompt sensitivity, retry depth, failure recovery strategy — translates directly into energy cost at the goal level\.EpGexposes the interaction between retry behavior and energy consumption\.

### 6\.1\.Goal\-Level Accounting

Energy\-per\-inference normalizes total system energy byNinferencesN\_\{\\mathrm\{inferences\}\}, a quantity the system designer controls\. By decomposing work into more numerous, cheaper calls, a system can report lower energy\-per\-inference while total goal energy rises\. The metric thereby becomes gameable by design\(Bertschet al\.,[2025](https://arxiv.org/html/2605.22883#bib.bib32); Kulkarni and others,[2026](https://arxiv.org/html/2605.22883#bib.bib31)\), and empirically it anti\-correlates with end\-to\-end task cost\(Oviedoet al\.,[2025](https://arxiv.org/html/2605.22883#bib.bib23)\)\.EpGsubstitutes the goal count, which is fixed by the evaluation protocol rather than by system design\. A benchmark defines how many goals are attempted; the system has no lever to adjust that denominator by changing its internal execution strategy\. The same goal\-level accountability applies symmetrically across architectures: a single\-shot linear system and a multi\-step agentic system solving identical tasks are compared under the same unit, with no structural advantage conferred to either by the metric itself\.

### 6\.2\.What Goal\-Level Accounting Enables

Consider two systems benchmarked on the same task family\. System A reports 50 J per inference; System B reports 70 J per inference\. Under inference\-level accounting, System A is the efficient choice\. At the goal level the picture inverts: System A requires on average four attempts before delivering a correct answer, accumulating 200 J per successful goal, while System B succeeds in one attempt at 70 J per goal\. The more reliable system consumes less energy per unit of useful work, a fact that inference\-level benchmarks not only fail to surface but actively obscure by rewarding low per\-call cost regardless of how many calls are needed\.EpGmakes this inversion visible and reproducible\. It enables direct comparison of systems that differ in reliability, architecture, and orchestration depth under a single unit that reflects what was actually delivered to the user\. The empirical analogue of this inversion, measured across 827 agentic and 827 linear goals on real hardware, is the central finding of Section[8](https://arxiv.org/html/2605.22883#S8)\.

### 6\.3\.Metric Properties and Boundary Conditions

The definition keeps success binary by design\. A goal either completes within the task definition’s acceptance criteria or it does not\. Binary success provides a clean metrological baseline: the denominator is unambiguous, reproducible across evaluators, and not subject to partial\-credit disagreements that would make cross\-paper comparison intractable\. For the empirical claims in this paper, binary success is both sufficient and appropriate: the tasks used are well\-specified arithmetic, logical, and factual problems with deterministic correct answers\.

One boundary condition requires handling\. When\|𝒲\+\|=0\|\\mathcal\{W\}^\{\+\}\|=0, the denominator is zero andEpGis undefined\. This arises when a system fails every goal in the evaluation set, a condition that is itself a meaningful experimental finding\. In such cases we report cumulative system energy alongside the observed success rate of 0%, rather than suppressing the result\. Latency and monetary cost are orthogonal performance dimensions and are reported alongsideEpGrather than collapsed into a composite figure, which would sacrifice the dimensional clarity that makes the unit scientifically useful\. All reportedEpGvalues are conditioned on the measurement boundaries declared in Section[3](https://arxiv.org/html/2605.22883#S3)and must be disclosed alongside any reported figure to enable meaningful cross\-study comparison\. A system measured with GPU energy included is not directly comparable to one measured at CPU package only; the boundary is part of the measurement, not a footnote to it\.

The empirical instantiation of this definition across 827 agentic and 827 linear goals yields meanEpGof 888\.1 J and 205\.3 J respectively, a ratio of 4\.33×\\timesthat forms the headline result of Section[8](https://arxiv.org/html/2605.22883#S8)\. The estimator is consistent and failure\-monotone under a truncated geometric retry model; the formal proofs, including closed\-form expectations and convergence analysis, are in Appendix[F](https://arxiv.org/html/2605.22883#A6)\.

## 7\.Orchestration Overhead Index \(OOI\)

EpGmeasures the absolute energy cost of completing a goal\. It does not answer the question a system designer must ask before deployment: is agentic execution structurally cheaper or more expensive than linear execution for this workload? Two systems with identicalEpGmay differ by an order of magnitude in orchestration overhead if one succeeds reliably on the first attempt and the other retries repeatedly\. Without a comparative unit, that overhead is invisible: every existing benchmark reports per\-inference energy or accuracy, leaving the goal\-level orchestration cost unmeasured, unreported, and unoptimized across the field\.OOIcloses this gap as a dimensionless complement toEpG, expressing orchestration overhead as a ratio that holds across hardware, models, and inference substrates\.

Formally, let𝒢\\mathcal\{G\}denote the set of evaluated goals and𝒲=\{agentic,linear\}\\mathcal\{W\}=\\\{\\textsc\{agentic\},\\textsc\{linear\}\\\}the set of workflow types\.OOIis a functionOOI:𝒢×𝒲2→ℝ\+\\textsc\{OOI\}\{\}:\\mathcal\{G\}\\times\\mathcal\{W\}^\{2\}\\rightarrow\\mathbb\{R\}^\{\+\}defined over matched pairs\(g,wa,wl\)\(g,w\_\{a\},w\_\{l\}\)where both workflow types attempt the same goalggunder identical task specification and success criterion\(Bertschet al\.,[2025](https://arxiv.org/html/2605.22883#bib.bib32); Afzalet al\.,[2023](https://arxiv.org/html/2605.22883#bib.bib33)\):

\(13\)OOI=EpGagenticEpGlinear\\textsc\{OOI\}\{\}=\\frac\{\\textsc\{EpG\}\_\{\\mathrm\{agentic\}\}\}\{\\textsc\{EpG\}\_\{\\mathrm\{linear\}\}\}
The range partitions into three metrologically distinct regimes\.OOI∈\(0,1\)\\textsc\{OOI\}\{\}\\in\(0,1\)identifies an orchestration dividend: agentic execution is strictly cheaper than linear per successful goal, arising when O\(1\) tool dispatch replaces token generation as the dominant cost path — agentic is the energy\-efficient choice\.OOI=1\\textsc\{OOI\}\{\}=1is the parity point: both workflows consume equal energy per goal\.OOI∈\(1,∞\)\\textsc\{OOI\}\{\}\\in\(1,\\infty\)quantifies the orchestration tax: linear execution is the energy\-efficient choice, with agentic overhead attributable to planning loops, inter\-step coordination, and retry amplification\. Two degenerate cases require explicit handling:OOI→∞\\textsc\{OOI\}\{\}\\rightarrow\\inftywhen the linear baseline succeeds but agentic never does, andOOI→0\\textsc\{OOI\}\{\}\\rightarrow 0in the converse; when both workflows fail,EpGis undefined for both andOOIis not computed — cumulative energy and success rate are reported separately per the boundary conditions of Section[6](https://arxiv.org/html/2605.22883#S6)\.

The linear baseline is the metrological anchor: a single prompt mapped to a single inference call, without tool use, branching, or retries\. It is the minimum\-energy path to goal completion under zero orchestration overhead\. Token volume is an experimental observation, not a control variable — the goal is defined by task specification and success criterion alone\. For workloads served by remote inference endpoints, the matched\-pair design controls for server\-side energy, making client\-sideOOIa lower bound on total orchestration overhead; full\-system scope is discussed in Section[10](https://arxiv.org/html/2605.22883#S10)\. BecauseEpGis consistent under the retry model of Section[6](https://arxiv.org/html/2605.22883#S6),OOIinherits consistency as a ratio of two convergent estimators\(Efron,[1979](https://arxiv.org/html/2605.22883#bib.bib60)\), stabilizing within the canonical goal set as shown in Figure[16](https://arxiv.org/html/2605.22883#A6.F16)\.

OOIenables three classes of decision thatEpGalone cannot support\. First,*task routing*: workload families withOOI<1\\textsc\{OOI\}\{\}<1are candidates for agentic execution by default, while families with highOOIrequire explicit cost justification against accuracy or latency gains\. Second,*model selection*: upgrading to a larger model may reduceEpGthrough fewer retries while increasing per\-inference cost —OOIreveals whether the net goal\-level effect is positive or negative\. Third,*fleet energy budgeting*: a mixed workload can be characterized by a portfolioOOIas the energy\-weighted mean across task families, enabling operators to set and enforce goal\-level energy SLAs independent of infrastructure substrate\. The empiricalOOIdistribution across all 827 canonical goals is reported in Section[8](https://arxiv.org/html/2605.22883#S8)and Figure[11](https://arxiv.org/html/2605.22883#S8.F11)\.

## 8\.Experimental Validation

This section validates theA\-LEMSmeasurement stack and derived metrics \(EpG,OOI\) through a structured causal elimination argument: each claim rules out one alternative explanation for the observed energy gap until orchestration structure alone remains\.

### 8\.1\.Evaluation Methodology

Platform\.A\-LEMSsamples IntelRAPLenergy counters\(David and others,[2010](https://arxiv.org/html/2605.22883#bib.bib9); Intel Corporation,[2023](https://arxiv.org/html/2605.22883#bib.bib10)\)at 100 Hz, applies2​σ2\\sigmaidle\-window baseline subtraction\(Thamm,[2025](https://arxiv.org/html/2605.22883#bib.bib27)\)to isolate workload\-induced energy from background system draw, and tracks execution within the attribution window\[t0,t1\]\[t\_\{0\},t\_\{1\}\]\(Section[3](https://arxiv.org/html/2605.22883#S3)\), which excludes pre\-task instrumentation and post\-task framework activity that prior tools conflate with workload energy\(Schmidt and others,[2022](https://arxiv.org/html/2605.22883#bib.bib5); Mattson and others,[2020](https://arxiv.org/html/2605.22883#bib.bib11)\)\. Every run is bound to its hardware and software context via the three\-hash protocolℋhw\\mathcal\{H\}\_\{\\mathrm\{hw\}\}/ℋenv\\mathcal\{H\}\_\{\\mathrm\{env\}\}/ℋrun\\mathcal\{H\}\_\{\\mathrm\{run\}\}\(Section[5](https://arxiv.org/html/2605.22883#S5)\)\. Full run configuration is given in Table[6](https://arxiv.org/html/2605.22883#S8.T6)\.

Two measurement regimes\.Two inference substrates are evaluated under the same measurement stack\.*Local inference*uses Ollama/TinyLlama\-1B \(n=588n=588\{\}agentic runs across all local experiments\), whereRAPLcaptures full package energy including all LLM computation\.*Remote inference*uses the Groq API with llama\-3\.3\-70b\-versatile \(n=378n=378\{\}agentic runs across all remote experiments\), whereRAPLcaptures only client\-side orchestration energy; server\-side computation is off\-host and not directly measured\. LLM API providers do not expose per\-request energy signals; the only published estimation method uses TDP scaled by PUE and wall\-clock time\(Patterson and others,[2021](https://arxiv.org/html/2605.22883#bib.bib2); Google Cloud,[2024](https://arxiv.org/html/2605.22883#bib.bib44)\), a linear proxy that does not capture per\-request or per\-phase variation and is therefore unsuitable for agentic workflow energy attribution\. Because both agentic and linear configurations invoke the same model with comparable task specifications, server\-side energy approximately cancels in theOOIratio, making client\-sideOOIa lower bound on total orchestration overhead\.

Task taxonomy\.Tasks span four structural tiers chosen to exercise distinct energy regimes \(Table[5](https://arxiv.org/html/2605.22883#S8.T5)\): factual retrieval \(factual QA, science QA\), mathematical reasoning \(GSM8K basic, GSM8K multi\-step\), logical reasoning, and tool\-augmented execution \(single\-tool and sequential chain\)\. This taxonomy follows the established classification of agentic workflow complexity\(Oviedoet al\.,[2025](https://arxiv.org/html/2605.22883#bib.bib23); Chienet al\.,[2023](https://arxiv.org/html/2605.22883#bib.bib25)\), ensuringOOIis evaluated across different execution structures rather than optimised for any particular outcome\. In tool\-augmented tasks, the linear workflow has no tool access and must resolve computation via LLM token generation; whetherOOI<1<1orOOI\>1\>1therefore tests whether the metric responds correctly to workflow structure\.

Table 5\.Task taxonomy and canonical results\. Evaluation functions are pre\-specified and fixed before execution\. OOI values from canonical paired set \(llama\_cpp, failure\_injection,session\_20260508\_190013\)\. Tool tasks use failure\_injection runs across all sessions\.†\\dagger=OOI<1\{<\}1: agentic more efficient than linear\.IDTierToolsPrompt TypeEvaluation FunctionSuccess CriterionOOIReasoning TasksFQAFactual retrievalnoneFactual questionExact string matchMatches reference answer4\.65×4\.65\\timesSciQAFactual retrievalnoneScience questionNormalized string match against accepted setAny reference matched5\.79×5\.79\\timesLRLogical reasoningnoneLogical syllogismExact label matchCorrect label selected4\.68×4\.68\\timesGSM8K\-BMath reasoningnoneArithmetic problemInteger exact match\(Cobbe and others,[2021](https://arxiv.org/html/2605.22883#bib.bib12)\)Ground\-truth integer2\.75×2\.75\\timesGSM8K\-MMath reasoningnoneMulti\-step arithmeticInteger exact match\(Cobbe and others,[2021](https://arxiv.org/html/2605.22883#bib.bib12)\)Ground\-truth integer7\.63×7\.63\\timesTool\-Augmented TasksTG:CalcSingle toolCalculatorFixed expressionDeterministic validatorExact numeric result0\.62×0\.62\\times†TG:DBSingle toolDatabaseFixed SQL queryDeterministic validatorExact query result0\.96×0\.96\\times†TG:Seq2Tool chainDB\+FileQuery then writeDeterministic validatorState matches spec1\.55×1\.55\\timesExperiment design\.Experiments are designed to isolate three distinct aspects of the thesis, each addressed by a dedicated study type\.*\(i\) Structural overhead study:*each goal is executed once as agentic and once as linear within the same session on the same goal instance, controlling for thermal state and DVFS drift via matchedℋhw\\mathcal\{H\}\_\{\\mathrm\{hw\}\}/ℋenv\\mathcal\{H\}\_\{\\mathrm\{env\}\}hashes; energy differences within a pair are attributable to workflow structure alone\. This yields827827\{\}matched agentic and827827\{\}linear goals across five reasoning task families andn=50n=50goal pairs per tool task family \(Table[5](https://arxiv.org/html/2605.22883#S8.T5)\)\.*\(ii\) Failure injection study:*to measure retry\-driven energy amplification, controlled failures are injected at a fixed rate\(Natellaet al\.,[2016](https://arxiv.org/html/2605.22883#bib.bib45)\), exercising the retry and recovery paths that production agentic systems traverse\(Jasperet al\.,[2025](https://arxiv.org/html/2605.22883#bib.bib46)\); this produces 851 total attempts, of which 29 retries following a failed first attempt\(3\.4% retry rate\)\.*\(iii\) Overhead study:*instrumentation overhead is measured across22282228\{\}runs from221221\{\}experiments spanning1111\{\}task families, confirming thatA\-LEMSdoes not contaminate reportedEpGvalues\. Together, these three studies provide the evidence base for claims C1–C5 in Sections[8\.2](https://arxiv.org/html/2605.22883#S8.SS2)–[8\.6](https://arxiv.org/html/2605.22883#S8.SS6)\.

### 8\.2\.C1 — Measurement Validity

TheA\-LEMSmeasurement stack produces valid energy measurements: the 100 Hz sampling target is empirically achieved,RAPLcounter deltas are uncorrupted, and sample coverage is sufficient for phase\-level attribution\.

Sampling rate\.The 100 Hz target \(Table[6](https://arxiv.org/html/2605.22883#S8.T6)\) holds in practice\. Across 4119580 samples from 588 local and 378 remote runs, the mean inter\-sample interval is 9\.71 ms \(103\.0 Hz\); 99\.85% of intervals fall within 5–15 ms and the maximum observed gap is 332\.6 ms, affecting<0\.01%<0\.01\\%of samples \(Figure[8](https://arxiv.org/html/2605.22883#S8.F8)\(a\)\)\. The distribution is identical between local and remote substrates, ruling out inference\-path effects on cadence\.

Counter reads\.Total run energy isE=RAPL​\(t1\)−RAPL​\(t0\)E=\\mathrm\{RAPL\}\(t\_\{1\}\)\-\\mathrm\{RAPL\}\(t\_\{0\}\)— two hardware counter reads whose accuracy is independent of sampling density \(Section[3](https://arxiv.org/html/2605.22883#S3)\)\. Across all 2228 runs, every delta is monotonic and non\-negative \(100\.0100\.0% L1 validity\): no wraparound, no missed reads\.EpGinherits this exactness directly\.

Coverage\.Samples serve phase attribution, not energy totals\. CoverageC=Tobserved/TtotalC=T\_\{\\mathrm\{observed\}\}/T\_\{\\mathrm\{total\}\}measures what fraction of\[t0,t1\]\[t\_\{0\},t\_\{1\}\]the sampler saw \(Section[4](https://arxiv.org/html/2605.22883#S4)\); short linear runs naturally score lower than long agentic runs, which motivates the gold thresholdC≥95%C\\geq 95\\%\(Figure[8](https://arxiv.org/html/2605.22883#S8.F8)\(b\)\)\. Of 2228 runs, 2006 reach gold tier, 140 are acceptable \(80≤C<95%80\\leq C<95\\%\), and 72 are excluded\. Across all five task families, mean coverage exceeds 90% for both workflow types \(Figure[8](https://arxiv.org/html/2605.22883#S8.F8)\(c\)\) — sufficient granularity to separate planning, execution, and synthesis phases\.

![Refer to caption](https://arxiv.org/html/2605.22883v1/x2.png)Figure 8\.\[C1\]\(a\) Inter\-sample interval distribution across 4119580 samples from both inference regimes: 99\.85% fall within 5–15 ms, confirming the 100 Hz target\. \(b\) Coverage vs\. run duration: short linear runs motivate the gold threshold \(C≥95%C\\geq 95\\%, dashed\)\. \(c\) Mean coverage by task and workflow type: all five canonical families exceed 90%, confirming phase attribution fidelity across the canonical dataset\.
### 8\.3\.C2 — Reproducibility

EveryA\-LEMSmeasurement is fully contextualized: hardware identity, software stack, and runtime state are recorded per run and queryable post\-hoc\.

Energy measurements move with hardware platform, OS and kernel, language runtime, CPU frequency governor, boost state, idle baseline, and thermal history\(Mytkowiczet al\.,[2009](https://arxiv.org/html/2605.22883#bib.bib47)\)\.A\-LEMSrecords or controls all seven \(Table[6](https://arxiv.org/html/2605.22883#S8.T6)\)\.

Across all 2228 runs, the CPU governor ispowersaveand turboenabled, verified against thebaseline\_idstored per run\. Mean idle package power across 48 baseline measurements is 2\.26 W; background CPU averages 2\.08% \(range 0\.6–5\.4%\), giving the2​σ2\\sigmasubtraction a stable floor\(Thamm,[2025](https://arxiv.org/html/2605.22883#bib.bib27)\)\. Thermal history is absorbed by the paired design: agentic and linear execute back\-to\-back on the same goal instance, so within\-pair energy differences reflect workflow structure\.

The three\-hash protocol \(Section[5](https://arxiv.org/html/2605.22883#S5)\) makes this context permanently queryable\. A replication attempt matchesℋhw=ebe694229b1b9d87\\mathcal\{H\}\_\{\\mathrm\{hw\}\}=\\texttt\{ebe694229b1b9d87\}and verifiesgovernor=powersave,turbo=enabledagainst the stored baseline record\. Table[4](https://arxiv.org/html/2605.22883#S5.T4)identifies what each hash mismatch indicates and which fields to inspect\.

Table 6\.Experimental configuration and context control across all 2228 runs\. All axes that independently affect energy measurement are either fixed by platform configuration or recorded per run for post\-hoc verification\.
### 8\.4\.C3 — Boundary Model Validation

Figure[9](https://arxiv.org/html/2605.22883#S8.F9)traces a representative paired run \(exp\. 629, GSM8K\-B, local llama\_cpp inference\) through the fourRAPLanchors that define theA\-LEMSmeasurement boundary\. Attpret\_\{\\mathrm\{pre\}\}, a counter snapshot establishes the pre\-task baseline window;t0t\_\{0\}marks the first energy sample and the start of attributed measurement;t1t\_\{1\}marks executor return and the end of attribution;t2t\_\{2\}closes the post\-task diagnostic window after framework teardown\. The agentic run draws623\.2623\.2J betweent0t\_\{0\}andt1t\_\{1\}; the linear run draws222\.6222\.6J over the same window\. Framework overhead — the energy consumed outside\[t0,t1\]\[t\_\{0\},t\_\{1\}\]— averages6\.8956\.895J for agentic workflows and4\.4244\.424J for linear workflows across4545normal comparison runs, representing1\.11\.1% and2\.122\.12% ofEpGrespectively\.

This fixed absolute overhead cost has a directional consequence for tools that estimate energy asTDP×twall\\mathrm\{TDP\}\\times t\_\{\\mathrm\{wall\}\}, wheretwallt\_\{\\mathrm\{wall\}\}includes post\-task activity\. Because linear workflows complete faster than agentic workflows for the same task, post\-task overhead constitutes a larger fraction of reported linear energy, compressing theOOIratio toward1\.0×1\.0\\timesindependently of any real difference in workload structure\. The\[t0,t1\]\[t\_\{0\},t\_\{1\}\]boundary eliminates this systematic bias:EpGis attributed exclusively to the executor window, and framework overhead is recorded separately as a platform characterization metric rather than conflated with workload energy\(Schmidt and others,[2022](https://arxiv.org/html/2605.22883#bib.bib5); Mattson and others,[2020](https://arxiv.org/html/2605.22883#bib.bib11)\)\.

The pre\-task diagnostic interval contributes87591\.087591\.0μ\\muJ on average — three orders of magnitude below task energy — and is excluded fromEpGby thet0t\_\{0\}anchor\. Post\-task energy, captured betweent1t\_\{1\}andt2t\_\{2\}, averages3389794\.03389794\.0μ\\muJ, reflecting the near\-idle power draw of framework teardown after the executor returns\. Neither window affects theOOIcomputation, which operates on attributed task energy only\.

![Refer to caption](https://arxiv.org/html/2605.22883v1/x3.png)Figure 9\.\[C3\]Measurement boundary trace for a representative paired run \(exp\. 629, GSM8K\-B, llama\_cpp, normal\)\. FourRAPLanchors partition execution into pre\-task, attributed task\[t0,t1\]\[t\_\{0\},t\_\{1\}\], and post\-task windows\. Framework overhead is1\.11\.1% of agenticEpGand2\.122\.12% of linearEpG— a fixed absolute cost that does not scale with task energy\. Tools estimating energy as TDP×\\timeswall\-time conflate this overhead with workload energy, compressingOOItoward1\.0×1\.0\\times\.
### 8\.5\.C4 — Discriminative Power

Across the canonical paired set of827827\{\}agentic and827827\{\}linear goals, agentic workflows consume a mean4\.33×4\.33\\timeshigher energy per goal \(888\.1888\.1J vs205\.3205\.3J\)\. TheOOIranges from3\.02×3\.02\\timesto7\.63×7\.63\\timesacross the five reasoning task families \(Figure[10](https://arxiv.org/html/2605.22883#S8.F10), Figure[11](https://arxiv.org/html/2605.22883#S8.F11)\), with higher values for tasks requiring more orchestration steps: GSM8K multi\-step \(7\.63×7\.63\\times\) involves more planning iterations than GSM8K basic \(2\.75×2\.75\\times\), and the overhead scales accordingly\. This correlation between task complexity andOOIis itself evidence that the metric captures orchestration depth, not a fixed measurement bias\. Contrary to our expectation, GSM8K\-M showed higher OOI than FQA — multi\-step arithmetic accumulated more retry overhead than factual retrieval across all canonical runs\.

![Refer to caption](https://arxiv.org/html/2605.22883v1/x4.png)Figure 10\.\[C4 Main Result\.\]\(a,b\) Local inference \(Ollama/TinyLlama\):EpGECDF and per\-taskOOIwith bootstrap 95% CIs\(500 resamples\)\. \(c,d\) Remote inference \(Groq/llama\-3\.3\-70b\): client\-sideEpGECDF and per\-taskOOI\.OOI\>1\>1in both regimes confirms that orchestration overhead is structural and substrate\-independent\.Tool\-task inversion\.Figure[11](https://arxiv.org/html/2605.22883#S8.F11)includes three tool\-graph task families alongside the reasoning tasks\. For single\-tool tasks \(tg\_single\_calc,tg\_single\_db\),OOI<1<1: agentic workflows are*more*energy\-efficient than their linear counterparts\. The mechanism is direct: the linear workflow has no tool access and must resolve the computation entirely via LLM token generation \- a billion\-parameter transformer reasoning through arithmetic token by token\. The agentic workflow dispatches a deterministic local tool and resolves the same computation inO​\(1\)O\(1\)\. Even with the orchestration loop overhead,EtotalE\_\{\\mathrm\{total\}\}is lower because the tool replaces the costliest part of the linear path\. For the sequential two\-tool chain,OOI=1\.55×=1\.55\\times— inter\-step coordination overhead partially offsets the tool efficiency gain\. These results confirm thatOOIis directionally correct across workflow structures and not biased upward by construction\. This finding also motivates workflow\-aware execution: for computation\-amenable tasks, agentic dispatch is the energy\-optimal choice\.

![Refer to caption](https://arxiv.org/html/2605.22883v1/x5.png)

Figure 11\.\[C4\+C5\]MeanEpGandOOIper task family\. Reasoning tasks \(top\) showOOI\>1\>1scaling with orchestration depth\. Tool tasks \(bottom\) showOOI≤1\\leq 1when tool execution replaces costlier LLM token generation\.OOIcorrectly captures the energy structure of each workflow type\.Task abbreviations follow Table[5](https://arxiv.org/html/2605.22883#S8.T5)\. T3:Hard \(n=19n=19\) is shown for completeness but excluded from the canonical paired set due to insufficient sample size\.

### 8\.6\.C5 — Orchestration Dominance

The tool\-task inversion confirmsOOIis not structurally biased upward\. The question is what drives the overhead in reasoning tasks\. Onn=305n=305\{\}goals with zero retry waste, agentic workflows consume1546\.01546\.0J versus linear315\.6315\.6J \(OOI=4\.9×=4\.9\\times, Figure[13](https://arxiv.org/html/2605.22883#S8.F13)\(b\)\)\. Retries are absent by construction: the entire gap originates in structural control flow — planning loops, multi\-step execution, and synthesis phases that linear workflows do not perform\. Orchestration structure is a sufficient cause of energy overhead, independent of retry behavior\. TheEpGgap between agentic and linear execution is driven by orchestration structure, not by increased computation\.

Retry amplification\.When failures do occur, energy amplifies further\. Across 851 total attempts in the retry study, 29 were retries; failed attempts consumed 26\.9% of total agentic energy \(Figure[12](https://arxiv.org/html/2605.22883#S8.F12), Figure[13](https://arxiv.org/html/2605.22883#S8.F13)\(a\)\)\. Each retry contributes one full attempt worth of energy to the numerator ofEpGwhile the denominator counts only successful goals, producing linear amplification with retry count\. Linear workflows achieve99\.899\.8% first\-attempt success, confirming that retry overhead is agentic\-specific\.

Substrate independence\.TheOOI\>1\>1result holds across both inference regimes\. Under local inference \(fullRAPL,n=588n=588\{\}runs\),OOIranges3\.02×3\.02\\times–7\.63×7\.63\\times\. Under remote inference \(client\-sideRAPLonly,n=378n=378\{\}runs\),OOI\>1\>1is confirmed across the same task families\. Because the two regimes use different hardware paths, different model sizes, and different energy measurement scopes, yet produce consistent directional results, the overhead cannot be attributed to the inference substrate — it originates in the orchestration layer\.

![Refer to caption](https://arxiv.org/html/2605.22883v1/x6.png)Figure 12\.\[C5 — Retry waste\.\]\(a\) MeanEpGlinear–agentic slope per task: all reasoning families show consistent agentic overhead\. \(b\) Useful \(green\) vs wasted \(red hatched\) energy per task: failed attempts account for 26\.9% of total agentic energy\.![Refer to caption](https://arxiv.org/html/2605.22883v1/x7.png)Figure 13\.\[C5 — Pure orchestration proof\.\]\(a\) Retry waste fraction per task: several task families show zero retry waste yet exhibitOOI\>1\>1in panel \(b\), confirming that retry amplification and structural control\-flow overhead are two independent mechanisms\. \(b\) Onn=305n=305\{\}goals with zero retry waste, agentic still consumes4\.9×4\.9\\timesmore energy than linear — structural orchestration overhead independent of retry behavior\.
### 8\.7\.Phase Power Characterisation

Table[7](https://arxiv.org/html/2605.22883#S8.T7)summarises per\-phase power and duration under local inference \(Ollama/TinyLlama,n=588n=588\{\}runs\)\. All three phases are measured and non\-zero, confirming full phase coverage with no external wait\. Figure[5](https://arxiv.org/html/2605.22883#S4.F5)\(c\) shows phase energy decomposition across canonical runs\.

Under remote inference \(Groq/llama\-3\.3\-70b,n=378n=378\{\}runs\), the local CPU draws0\.20\.2W during the planning phase yet averages1\.01\.0W across the full task duration \(mean API wait:14913\.014913\.0ms\), with the gap attributable to orchestration activity during remote token generation\. The elevated overall power reveals that the orchestration framework maintains non\-trivial local CPU activity during remote token generation — coordination, state management, and response buffering that inference\-level benchmarks\(Mattson and others,[2020](https://arxiv.org/html/2605.22883#bib.bib11)\)cannot observe because they do not instrument the client\-side orchestration layer\. This local orchestration draw is energy that standard benchmarks would misattribute to inference compute; our phase\-resolved measurement captures it explicitly\.

Table 7\.Per\-phase energy, power, and duration: agentic workflows, local inference \(Ollama/TinyLlama,n=588n=588\{\}, fullRAPLcompute\) vs\. remote inference \(Groq/llama\-3\.3\-70b,n=378n=378\{\}, client\-side orchestration only\)\.
### 8\.8\.Measurement Scope

A\-LEMSmeasures local CPU energy viaRAPL\. GPU energy, network interface energy, and remote server\-side computation are not directly measured\. For remote inferenceOOI, server\-side energy approximately cancels in the ratio since both workflows invoke the same remote model\. The cancellation is not exact: agentic workflows trigger additional server\-side calls during planning and retry phases that linear workflows do not, making client\-sideOOI=4\.33×=4\.33\\timesa lower bound on the fully\-attributed remote overhead\. Local inferenceOOIis not subject to this limitation:RAPLcaptures full package energy including all LLM computation\. The canonical result is obtained on a single hardware platform \(ℋhw=\\mathcal\{H\}\_\{\\mathrm\{hw\}\}=ebe694229b1b9d87\) using TinyLlama\-1B for local inference; generalization to larger models and multi\-platform deployments is future work\. TheOOI\>1\>1result under remote inference \(Groq/llama\-3\.3\-70b,n=378n=378\{\}runs\) — where localRAPLcaptures only orchestration\-layer activity with no local LLM computation — provides independent evidence that the overhead originates in orchestration, not model size\. A\-LEMS validates the measurement framework on a single hardware\-grounded instance; the thesis is unit necessity, not population characterization — a single real system suffices to demonstrate that inference\-level accounting fails and goal\-level accounting does not\.

## 9\.Related Work

### 9\.1\.Inference\-Level Energy Reporting

Strubell et al\.\(Strubellet al\.,[2019](https://arxiv.org/html/2605.22883#bib.bib1)\)and Patterson et al\.\(Patterson and others,[2021](https://arxiv.org/html/2605.22883#bib.bib2)\)establish foundational work on energy and carbon reporting for large\-scale neural network training and inference\. MLPerf Power\(Tschand and others,[2024](https://arxiv.org/html/2605.22883#bib.bib51)\)extends this to a comprehensive benchmarking methodology spanning 60 systems and 1,841 reproducible measurements across µW\-to\-MW deployment scales, establishing energy efficiency as a first\-class metric in ML system evaluation\. These efforts share a common assumption: inference is the atomic unit of measurement\. For classical single\-pass workloads this holds\. In agentic systems, where a single user goal triggers planning loops, tool dispatch, retries, and recovery cycles whose depth is determined at runtime, the inference boundary no longer coincides with the task boundary\. Counting inferences measures implementation behavior, not goal completion cost\.

Theml\.energybenchmark\(Chunget al\.,[2025](https://arxiv.org/html/2605.22883#bib.bib62)\)automates inference energy measurement with service\-aware accounting across 40 model architectures, normalizing by request — coherent for single\-pass serving workloads — but does not extend to multi\-step agentic execution where request count is determined at runtime rather than by task definition\.

### 9\.2\.Hardware\-Level Energy Characterization

Proxy\-based estimation using FLOPs or thermal design power \(TDP\) remains common but cannot distinguish computation from orchestration overhead or attribute cost across execution phases\. Patel et al\.\(Patelet al\.,[2024](https://arxiv.org/html/2605.22883#bib.bib50)\)take a direct measurement approach, extensively characterizing GPU power consumption patterns for LLM training and inference at datacenter scale, showing that inference clusters carry substantial headroom for power oversubscription and proposing POLCA to exploit it\. Their work operates at infrastructure granularity, optimizing power allocation assuming inference as the atomic unit\.A\-LEMSoperates at a finer granularity: it attributes energy to individual workflow phases and successful goals, making orchestration overhead visible where infrastructure\-level tools cannot\.

### 9\.3\.System\-Level Power Profiling Tools

PowerAPI\(Bourdon and others,[2013](https://arxiv.org/html/2605.22883#bib.bib7)\), Scaphandre\(Petit and others,[2021](https://arxiv.org/html/2605.22883#bib.bib6)\), and CodeCarbon\(Schmidt and others,[2022](https://arxiv.org/html/2605.22883#bib.bib5)\)provide system\-level power measurement infrastructure thatA\-LEMSbuilds upon at the signal layer\. None defines a goal\-level energy unit, attributes cost across orchestration phases, or accounts for retry\-driven energy amplification\(Table[8](https://arxiv.org/html/2605.22883#S9.T8)\)\.EpGis constructed on top of these signals with an explicit attribution boundary and a goal\-level denominator that these tools do not provide\.

### 9\.4\.Agentic AI System Characterization

Raj et al\.\(Rajet al\.,[2025](https://arxiv.org/html/2605.22883#bib.bib49)\)characterize agentic AI execution from a CPU\-centric perspective, identifying orchestration and tool dispatch as dominant sources of latency and CPU utilization on heterogeneous CPU\-GPU systems\. Their scheduling optimizations reduce end\-to\-end latency by up to 3\.9×\\times\. Where their work optimizes latency under fixed inference cost assumptions,A\-LEMSmeasures the energy cost of orchestration structure itself: the overhead that persists even after scheduling is optimized\.

Chen et al\.\(Chen and others,[2026](https://arxiv.org/html/2605.22883#bib.bib17)\)survey networking\-aware energy efficiency in distributed agentic inference, proposing a taxonomy across model simplification, computation control, and cross\-layer co\-design\. The survey establishes that computation and communication costs compound across multi\-step pipelines, but does not provide a hardware\-grounded unit for attributing energy to individual orchestration decisions\.EpGfills this gap by measuring goal\-level energy directly from RAPL counters under a tightly defined attribution boundary\.

Table 8\.Comparative analysis of energy measurement systems\.✓= supported;∼\\sim= partial;×\\times= not supported;−\-= not applicable\.A\-LEMSis the only system providing goal\-level energy accounting, explicit attribution boundaries, retry energy capture, and a reproducibility protocol for agentic AI workloads\.Dimension

A\-LEMS\(ours\)

PowerAPI\(Bourdon and others,[2013](https://arxiv.org/html/2605.22883#bib.bib7)\)

Scaphandre\(Petit and others,[2021](https://arxiv.org/html/2605.22883#bib.bib6)\)

CodeCarbon\(Schmidt and others,[2022](https://arxiv.org/html/2605.22883#bib.bib5)\)

MLPerf Power\(Tschand and others,[2024](https://arxiv.org/html/2605.22883#bib.bib51)\)

ml\.energy\(Chunget al\.,[2025](https://arxiv.org/html/2605.22883#bib.bib62)\)

Raj et al\.\(Rajet al\.,[2025](https://arxiv.org/html/2605.22883#bib.bib49)\)

Chen et al\.\(Chen and others,[2026](https://arxiv.org/html/2605.22883#bib.bib17)\)

Energy UnitGoal\-level unit \(J/goal\)✓×\\times×\\times×\\times×\\times×\\times×\\times×\\timesRetry energy captured✓×\\times×\\times×\\times×\\times×\\times×\\times×\\timesOrchestration ratio \(OOI\)✓×\\times×\\times×\\times×\\times×\\times×\\times×\\timesMeasurementHardware measurement✓✓✓∼\\sim✓✓∼\\sim×\\timesAttribution boundary✓×\\times×\\times×\\times∼\\sim×\\times×\\times×\\timesPhase\-level attribution✓×\\times×\\times×\\times×\\times∼\\sim✓×\\timesBaseline subtraction✓∼\\sim∼\\sim×\\times∼\\sim∼\\sim×\\times×\\timesMethodologyReproducibility protocol✓×\\times×\\times×\\times✓✓×\\times×\\timesAgentic workflow support✓×\\times×\\times×\\times×\\times×\\times✓∼\\simEmpirical validation✓×\\times×\\times×\\times✓✓✓×\\times
### 9\.5\.PUE as Historical Precedent

PUE\(Barroso and Hölzle,[2007](https://arxiv.org/html/2605.22883#bib.bib13); The Green Grid,[2012](https://arxiv.org/html/2605.22883#bib.bib14)\)succeeded not because it is theoretically complete but because it is operationally defined, empirically measurable, and names its gaming failure modes explicitly\. A datacenter can game PUE by running servers at full load during measurement; the metric acknowledges this and specifies conditions under which comparisons are valid\.EpGfollows the same design philosophy: it defines its attribution window precisely, binds results to hardware and environment state, and documents the boundary failure modes that would corrupt comparison across systems\.

## 10\.Discussion, Limitations, and Future Directions

### 10\.1\.Limitations

EpGmeasures energy per successful goal under a tightly defined attribution boundary, but three limitations bound its current scope\. First, success is binary:EpGdoes not capture gradations in output quality, so a system that produces marginally correct answers at low energy looks identical to one that produces high\-quality answers at the same cost\. Second, the measurement substrate is local CPU viaRAPL\. GPU energy, network interface energy, and remote server\-side computation are not directly measured\. For reasoning tasks under local inference this is sufficient; for remote inference deployments,EpGcaptures orchestration\-layer cost only, and full system attribution requires provider\-disclosed server\-side energy data that is not yet standardized\(Patelet al\.,[2024](https://arxiv.org/html/2605.22883#bib.bib50)\)\. Third,OOIis a comparative ratio requiring a matched linear baseline — it cannot be reported for a system in isolation without defining the linear reference point\.

### 10\.2\.Metric Integrity

Any efficiency metric creates incentives for gaming, andEpGis no exception\. A system operator could bias task selection toward easier goals to reduce reportedEpG, terminate execution early to reduce energy accumulation while preserving apparent success rate, or shift measurement boundaries to exclude expensive initialization phases\. The three\-hash reproducibility protocol mitigates boundary shifting by binding every run to its exact hardware and software context\. Task distribution gaming is mitigated by reporting success rates and task family breakdowns alongsideEpG\(Pineauet al\.,[2021](https://arxiv.org/html/2605.22883#bib.bib48)\)— a system that achieves lowEpGon a restricted easy\-task subset will show a different task distribution than one evaluated on the full benchmark\. Early termination is detectable through completion statistics\. These mitigations follow the same design philosophy as PUE: name the gaming modes and specify the measurement conditions under which comparisons are valid\(Barroso and Hölzle,[2007](https://arxiv.org/html/2605.22883#bib.bib13)\)\.

### 10\.3\.Future Directions

Local orchestration overhead is a measurable, substantial fraction of local CPU energy in agentic workflows, and quantifying it is necessary, though not sufficient, for full system energy accounting\. Extending the attribution substrate beyond local CPU to GPU, network, and server\-side components remains an open problem whose solution would enable complete cross\-layer energy accounting\(Rajet al\.,[2025](https://arxiv.org/html/2605.22883#bib.bib49); Patelet al\.,[2024](https://arxiv.org/html/2605.22883#bib.bib50)\)\.

### 10\.4\.Conclusion

Energy\-per\-inference is the wrong unit for agentic AI\. It normalizes by implementation steps rather than goal completions, omits retry and recovery energy by construction, and cannot distinguish a system that succeeds on the first attempt from one that fails four times before succeeding\. These are not edge cases — they define the normal operating behavior of production agentic systems\(Jasperet al\.,[2025](https://arxiv.org/html/2605.22883#bib.bib46); Cemriet al\.,[2025](https://arxiv.org/html/2605.22883#bib.bib61)\)\.

EpG\(Energy per Successful Goal, J/goal\) corrects this by aggregating total workflow energy across all attempts and normalizing by successfully completed goals, grounded in hardware\-levelRAPLmeasurement with explicit attribution boundaries and a three\-hash reproducibility protocol\.OOIisolates the energy cost of orchestration structure relative to matched linear execution, revealing a4\.33×4\.33\\timesoverhead in canonical paired runs — driven not by increased inference compute but by planning loops, retry cycles, and coordination overhead that linear workflows do not perform\. The tool\-task inversion \(OOI<1<1for single\-tool dispatch\) confirms the metric is directionally correct: it measures orchestration structure, not a fixed upward bias\.

Optimizing inference efficiency without accounting for orchestration structure addresses the wrong layer\.EpGandOOIgive system designers and benchmark authors a hardware\-grounded unit to measure and reduce the true energy cost of agentic AI workloads\.

## References

- A\. Afzal, G\. Hager, and G\. Wellein \(2023\)SPEChpc 2021 benchmarks on ice lake and sapphire rapids infiniband clusters: a performance and energy case study\.InProceedings of the ACM International Conference on High Performance Computing, Networking, Storage and Analysis,External Links:[Document](https://dx.doi.org/10.1145/3624062.3624197),[Link](https://dl.acm.org/doi/10.1145/3624062.3624197)Cited by:[§6](https://arxiv.org/html/2605.22883#S6.p1.7),[§7](https://arxiv.org/html/2605.22883#S7.p2.5)\.
- Allianz Research \(2026\)Thinking fast, building slow: the energy cost of the US AI boom\.Allianz Economic Research\.Cited by:[§1\.5](https://arxiv.org/html/2605.22883#S1.SS5.p1.1),[§1\.5](https://arxiv.org/html/2605.22883#S1.SS5.p3.1)\.
- S\. Aquino\-Brítez, P\. García\-Sánchez, A\. Ortiz, and D\. Aquino\-Brítez \(2025\)Towards an energy consumption index for deep learning models: a comparative analysis of architectures, GPUs, and measurement tools\.Sensors25\(3\),pp\. 846\.External Links:[Document](https://dx.doi.org/10.3390/s25030846),[Link](https://doi.org/10.3390/s25030846)Cited by:[§4\.2](https://arxiv.org/html/2605.22883#S4.SS2.p1.1)\.
- L\. A\. Barroso and U\. Hölzle \(2007\)The case for energy\-proportional computing\.IEEE Computer40\(12\),pp\. 33–37\.Note:Paper 1: §1 PUE analogy, §8 positioning\.Cited by:[§1\.6](https://arxiv.org/html/2605.22883#S1.SS6.p1.1),[§10\.2](https://arxiv.org/html/2605.22883#S10.SS2.p1.1),[Table 1](https://arxiv.org/html/2605.22883#S3.T1.12.10.5.1.1),[§4\.2](https://arxiv.org/html/2605.22883#S4.SS2.p1.1),[§9\.5](https://arxiv.org/html/2605.22883#S9.SS5.p1.1)\.
- A\. Bertsch, M\. R\. Collette, S\. A\. Dawson, S\. D\. Hammond, I\. Karlin, M\. S\. McKinley, K\. Pedretti, R\. N\. Rieben, B\. S\. Ryujin, A\. Vargas, and K\. Weiss \(2025\)Understanding power and energy utilization in large scale production physics simulation codes\.The International Journal of High Performance Computing Applications\.External Links:[Document](https://dx.doi.org/10.1177/1094342025136263),[Link](https://doi.org/10.1177/1094342025136263)Cited by:[§4\.2](https://arxiv.org/html/2605.22883#S4.SS2.p1.1),[§6\.1](https://arxiv.org/html/2605.22883#S6.SS1.p1.1),[§6](https://arxiv.org/html/2605.22883#S6.p1.7),[§7](https://arxiv.org/html/2605.22883#S7.p2.5)\.
- A\. Bourdonet al\.\(2013\)PowerAPI: a software library to monitor the energy consumed at the process\-level\.ERCIM News\(92\)\.Note:Paper 2: §2 comparison table\.Cited by:[§9\.3](https://arxiv.org/html/2605.22883#S9.SS3.p1.1),[Table 8](https://arxiv.org/html/2605.22883#S9.T8.65.60.1.3.1.1.1)\.
- M\. Cemri, M\. Z\. Pan, S\. Yang,et al\.\(2025\)Why do multi\-agent LLM systems fail?\.arXiv preprint arXiv:2503\.13657\.External Links:[Link](https://arxiv.org/abs/2503.13657)Cited by:[§10\.4](https://arxiv.org/html/2605.22883#S10.SS4.p1.1)\.
- X\. Chenet al\.\(2026\)Networking\-aware energy efficiency in agentic AI inference: a survey\.arXiv preprint 2604\.07857\.Note:Paper 1: §8 MUST CITE — same problem space, survey not empirical\. Paper 2: §2 acknowledge\. Paper 3: §9 positioning\.Cited by:[§1\.5](https://arxiv.org/html/2605.22883#S1.SS5.p1.1),[§9\.4](https://arxiv.org/html/2605.22883#S9.SS4.p2.1),[Table 8](https://arxiv.org/html/2605.22883#S9.T8.65.60.1.9.1.1.1)\.
- A\. A\. Chien, L\. Lin, H\. Nguyen, V\. Rao, T\. Sharma, and R\. Wijayawardana \(2023\)Reducing the carbon impact of generative AI inference \(today and in 2035\)\.InProceedings of the 2nd Workshop on Sustainable Computer Systems Design and Implementation \(HotCarbon ’23\),External Links:[Document](https://dx.doi.org/10.1145/3604930.3605705),[Link](https://dl.acm.org/doi/10.1145/3604930.3605705)Cited by:[§1\.1](https://arxiv.org/html/2605.22883#S1.SS1.p2.1),[§1\.6](https://arxiv.org/html/2605.22883#S1.SS6.p2.1),[§8\.1](https://arxiv.org/html/2605.22883#S8.SS1.p3.2)\.
- J\. Chung, J\. J\. Ma, R\. Wu, J\. Liu, O\. J\. Kweon, Y\. Xia, Z\. Wu, and M\. Chowdhury \(2025\)The ML\.ENERGY benchmark: toward automated inference energy measurement and optimization\.InNeurIPS Datasets and Benchmarks,Cited by:[§1\.1](https://arxiv.org/html/2605.22883#S1.SS1.p1.1),[§9\.1](https://arxiv.org/html/2605.22883#S9.SS1.p2.1),[Table 8](https://arxiv.org/html/2605.22883#S9.T8.65.60.1.7.1.1.1.1)\.
- K\. Cobbeet al\.\(2021\)Training verifiers to solve math word problems\.arXiv preprint 2110\.14168\.Note:All 3 papers: GSM8K task reference\.Cited by:[§2\.3](https://arxiv.org/html/2605.22883#S2.SS3.p3.1),[Table 5](https://arxiv.org/html/2605.22883#S8.T5.8.4.6.1.1),[Table 5](https://arxiv.org/html/2605.22883#S8.T5.9.5.6.1.1)\.
- H\. Davidet al\.\(2010\)RAPL: memory power estimation and capping\.InProc\. ISLPED,pp\. 189–194\.Note:All 3 papers: RAPL reference\.Cited by:[§4\.1](https://arxiv.org/html/2605.22883#S4.SS1.p1.1),[§8\.1](https://arxiv.org/html/2605.22883#S8.SS1.p1.5)\.
- B\. Efron \(1979\)Bootstrap methods: another look at the jackknife\.The Annals of Statistics7\(1\),pp\. 1–26\.Cited by:[§7](https://arxiv.org/html/2605.22883#S7.p5.1)\.
- European Parliament and Council of the European Union \(2024\)Regulation \(EU\) 2024/1689 on artificial intelligence \(EU AI Act\)\.Note:Official Journal of the European UnionCited by:[§1\.5](https://arxiv.org/html/2605.22883#S1.SS5.p2.1)\.
- European Parliamentary Research Service \(2025\)AI and the energy sector\.Technical reportEuropean Parliament\.Cited by:[§1\.5](https://arxiv.org/html/2605.22883#S1.SS5.p2.1)\.
- Google Cloud \(2024\)Carbon footprint methodology\.Note:[https://cloud\.google\.com/carbon\-footprint/docs/methodology](https://cloud.google.com/carbon-footprint/docs/methodology)Accessed: 2026\-05\-10Cited by:[§1\.4](https://arxiv.org/html/2605.22883#S1.SS4.p3.2),[Table 1](https://arxiv.org/html/2605.22883#S3.T1.6.4.1.1.1),[§8\.1](https://arxiv.org/html/2605.22883#S8.SS1.p2.2)\.
- Intel Corporation \(2023\)Intel 64 and IA\-32 architectures software developer’s manual, vol\. 3b: RAPL interface\.Note:Technical ReportAll 3 papers: RAPL specification\.Cited by:[§4\.1](https://arxiv.org/html/2605.22883#S4.SS1.p1.1),[§8\.1](https://arxiv.org/html/2605.22883#S8.SS1.p1.5)\.
- S\. Jasper, M\. Luu, E\. Pan, A\. Tyagi, M\. Quinn, J\. Hu, and D\. Houngninou \(2025\)BugGen: a self\-correcting multi\-agent LLM pipeline for realistic RTL bug synthesis\.InACM/IEEE Symposium on Machine Learning for CAD \(MLCAD\),pp\. 1–9\.Cited by:[§10\.4](https://arxiv.org/html/2605.22883#S10.SS4.p1.1),[§8\.1](https://arxiv.org/html/2605.22883#S8.SS1.p4.8)\.
- N\. Jegham, M\. Abdelatti, C\. Y\. Koh, L\. Elmoubarki, and A\. Hendawi \(2025\)How hungry is AI? benchmarking energy, water, and carbon footprint of LLM inference\.arXiv preprint arXiv:2505\.09598\.External Links:[Document](https://dx.doi.org/10.48550/arXiv.2505.09598),[Link](https://arxiv.org/abs/2505.09598)Cited by:[§1\.1](https://arxiv.org/html/2605.22883#S1.SS1.p1.1),[§1\.3](https://arxiv.org/html/2605.22883#S1.SS3.p2.2),[§1\.5](https://arxiv.org/html/2605.22883#S1.SS5.p1.1),[§1\.6](https://arxiv.org/html/2605.22883#S1.SS6.p2.1)\.
- T\. Kneese and M\. Young \(2024\)Carbon emissions in the tailpipe of generative AI\.Harvard Data Science Review6\(S5\)\.External Links:[Document](https://dx.doi.org/10.1162/99608f92.fbdf6128),[Link](https://doi.org/10.1162/99608f92.fbdf6128)Cited by:[§1\.3](https://arxiv.org/html/2605.22883#S1.SS3.p2.2)\.
- G\. Koszczał, M\. Matuszek, and P\. Czarnul \(2025\)Comparison and analysis of software and hardware energy measurement methods for a CPU\+GPU system and selected parallel applications\.Computer Science and Information Systems22\(2\),pp\. 563–590\.External Links:[Document](https://dx.doi.org/10.2298/CSIS240722023K),[Link](https://doi.org/10.2298/CSIS240722023K)Cited by:[§4\.1](https://arxiv.org/html/2605.22883#S4.SS1.p1.1),[§4\.1](https://arxiv.org/html/2605.22883#S4.SS1.p3.1)\.
- K\. Kulkarniet al\.\(2026\)Harvesting energy consumption on european HPC systems: sharing experience from the CEEC project\.InProceedings of Supercomputing Asia and International Conference on High Performance Computing in Asia\-Pacific Region Workshops,External Links:[Document](https://dx.doi.org/10.1145/3784828.3785161),[Link](https://doi.org/10.1145/3784828.3785161)Cited by:[§4\.1](https://arxiv.org/html/2605.22883#S4.SS1.p3.1),[§6\.1](https://arxiv.org/html/2605.22883#S6.SS1.p1.1),[§6](https://arxiv.org/html/2605.22883#S6.p1.7)\.
- R\. Mahapatra and W\. Zhao \(2005\)An energy\-efficient slack distribution technique for multimode distributed real\-time embedded systems\.Parallel and Distributed Systems, IEEE Transactions on16,pp\. 650– 662\.External Links:[Document](https://dx.doi.org/10.1109/TPDS.2005.78)Cited by:[Table 1](https://arxiv.org/html/2605.22883#S3.T1.12.10.5.1.1)\.
- P\. Mattsonet al\.\(2020\)MLPerf training benchmark\.InProc\. MLSys,Note:Paper 1: §8 related work \(throughput benchmark\)\. Paper 3: §9 positioning\.Cited by:[§3\.3](https://arxiv.org/html/2605.22883#S3.SS3.p2.1),[Table 1](https://arxiv.org/html/2605.22883#S3.T1.5.3.6.1.1),[§8\.1](https://arxiv.org/html/2605.22883#S8.SS1.p1.5),[§8\.4](https://arxiv.org/html/2605.22883#S8.SS4.p2.4),[§8\.7](https://arxiv.org/html/2605.22883#S8.SS7.p2.4)\.
- T\. Mytkowicz, A\. Diwan, M\. Hauswirth, and P\. F\. Sweeney \(2009\)Producing wrong data without doing anything obviously wrong\!\.InProceedings of the 14th International Conference on Architectural Support for Programming Languages and Operating Systems \(ASPLOS\),pp\. 265–276\.Cited by:[§1\.4](https://arxiv.org/html/2605.22883#S1.SS4.p5.1),[§5](https://arxiv.org/html/2605.22883#S5.p1.1),[§8\.3](https://arxiv.org/html/2605.22883#S8.SS3.p2.1)\.
- R\. Natella, D\. Cotroneo, and H\. S\. Madeira \(2016\)Assessing dependability with software fault injection: a survey\.ACM Computing Surveys48\(3\)\.External Links:[Document](https://dx.doi.org/10.1145/2841425)Cited by:[§8\.1](https://arxiv.org/html/2605.22883#S8.SS1.p4.8)\.
- H\. A\. Niaz, R\. R\. Manumachu, and A\. L\. Lastovetsky \(2025\)Accurate and reliable energy measurement and modelling of data transfer between CPU and GPU in parallel applications on heterogeneous hybrid platforms\.IEEE Transactions on Computers74\(3\),pp\. 1011–1024\.External Links:[Document](https://dx.doi.org/10.1109/TC.2024.3504262),[Link](https://doi.org/10.1109/TC.2024.3504262)Cited by:[§4\.1](https://arxiv.org/html/2605.22883#S4.SS1.p3.1)\.
- F\. Oviedo, F\. Kazhamiaka, E\. Choukse, A\. Kim, A\. Luers, M\. Nakagawa, R\. Bianchini, and J\. M\. Lavista Ferres \(2025\)Energy use of AI inference: efficiency pathways and test\-time compute\.arXiv preprint arXiv:2509\.20241\.External Links:[Document](https://dx.doi.org/10.48550/arXiv.2509.20241),[Link](https://arxiv.org/abs/2509.20241)Cited by:[§1\.1](https://arxiv.org/html/2605.22883#S1.SS1.p1.1),[§1\.1](https://arxiv.org/html/2605.22883#S1.SS1.p2.1),[§1\.5](https://arxiv.org/html/2605.22883#S1.SS5.p1.1),[§6\.1](https://arxiv.org/html/2605.22883#S6.SS1.p1.1),[§8\.1](https://arxiv.org/html/2605.22883#S8.SS1.p3.2)\.
- P\. Patel, E\. Choukse, C\. Zhang, Í\. Goiri, B\. Warrier, N\. Mahalingam, and R\. Bianchini \(2024\)Characterizing power management opportunities for LLMs in the cloud\.InProceedings of the 29th ACM International Conference on Architectural Support for Programming Languages and Operating Systems,ASPLOS ’24\.Cited by:[§10\.1](https://arxiv.org/html/2605.22883#S10.SS1.p1.1),[§10\.3](https://arxiv.org/html/2605.22883#S10.SS3.p1.1),[§9\.2](https://arxiv.org/html/2605.22883#S9.SS2.p1.1)\.
- D\. Pattersonet al\.\(2021\)Carbon emissions and large neural network training\.arXiv preprint 2104\.10350\.Note:Paper 1: §1 motivation\. Paper 2: §1 motivation\.Cited by:[§1\.1](https://arxiv.org/html/2605.22883#S1.SS1.p1.1),[§8\.1](https://arxiv.org/html/2605.22883#S8.SS1.p2.2),[§9\.1](https://arxiv.org/html/2605.22883#S9.SS1.p1.1)\.
- B\. Petitet al\.\(2021\)Scaphandre: a metrology agent dedicated to measure the energy consumption of IT services\.InIEEE MASCOTS,Note:Paper 2: §2 comparison table\.Cited by:[Table 1](https://arxiv.org/html/2605.22883#S3.T1.10.8.4.1.1),[§9\.3](https://arxiv.org/html/2605.22883#S9.SS3.p1.1),[Table 8](https://arxiv.org/html/2605.22883#S9.T8.65.60.1.4.1.1.1)\.
- J\. Pineau, P\. Vincent\-Lamarre, K\. Sinha, V\. Larivière, A\. Beygelzimer, F\. d’Alché\-Buc, E\. Fox, and H\. Larochelle \(2021\)Improving reproducibility in machine learning research\.Journal of Machine Learning Research22\(164\),pp\. 1–20\.Cited by:[§1\.4](https://arxiv.org/html/2605.22883#S1.SS4.p5.1),[§10\.2](https://arxiv.org/html/2605.22883#S10.SS2.p1.1)\.
- R\. Raj, S\. Kundu, I\. Vohra, H\. Wang, and T\. Krishna \(2025\)Towards understanding, analyzing, and optimizing agentic AI execution: a CPU\-centric perspective\.External Links:2511\.00739Cited by:[§10\.3](https://arxiv.org/html/2605.22883#S10.SS3.p1.1),[§9\.4](https://arxiv.org/html/2605.22883#S9.SS4.p1.1),[Table 8](https://arxiv.org/html/2605.22883#S9.T8.65.60.1.8.1.1.1)\.
- V\. Schmidtet al\.\(2022\)CodeCarbon: estimate and track carbon emissions from machine learning\.arXiv preprint 2002\.05651\.Note:Paper 1: §3 boundary failure mode \(inflation\)\. Paper 2: §2 comparison table\.Cited by:[§1\.4](https://arxiv.org/html/2605.22883#S1.SS4.p3.2),[§3\.3](https://arxiv.org/html/2605.22883#S3.SS3.p2.1),[Table 1](https://arxiv.org/html/2605.22883#S3.T1.6.4.1.1.1),[§8\.1](https://arxiv.org/html/2605.22883#S8.SS1.p1.5),[§8\.4](https://arxiv.org/html/2605.22883#S8.SS4.p2.4),[§9\.3](https://arxiv.org/html/2605.22883#S9.SS3.p1.1),[Table 8](https://arxiv.org/html/2605.22883#S9.T8.65.60.1.5.1.1.1)\.
- Schneider Electric \(2025\)The unseen cost of artificial intelligence: energy and water consumption\.Technical reportSchneider Electric\.Cited by:[§1\.5](https://arxiv.org/html/2605.22883#S1.SS5.p1.1),[§1\.5](https://arxiv.org/html/2605.22883#S1.SS5.p2.1)\.
- E\. Strubell, A\. Ganesh, and A\. McCallum \(2019\)Energy and policy considerations for deep learning in NLP\.InProc\. ACL,pp\. 3645–3650\.Note:Paper 1: §1 motivation\. Paper 2: §1 motivation\.Cited by:[§1\.1](https://arxiv.org/html/2605.22883#S1.SS1.p1.1),[§9\.1](https://arxiv.org/html/2605.22883#S9.SS1.p1.1)\.
- P\. Thamm \(2025\)Strategies to measure energy consumption using RAPL during workflow execution on commodity clusters\.arXiv preprint arXiv:2505\.09375\.External Links:[Document](https://dx.doi.org/10.48550/arXiv.2505.09375),[Link](https://arxiv.org/abs/2505.09375)Cited by:[§1\.4](https://arxiv.org/html/2605.22883#S1.SS4.p4.1),[§3\.3](https://arxiv.org/html/2605.22883#S3.SS3.p2.1),[§4\.1](https://arxiv.org/html/2605.22883#S4.SS1.p1.1),[§5](https://arxiv.org/html/2605.22883#S5.p1.1),[§8\.1](https://arxiv.org/html/2605.22883#S8.SS1.p1.5),[§8\.3](https://arxiv.org/html/2605.22883#S8.SS3.p3.1)\.
- The Green Grid \(2012\)Power usage effectiveness \(PUE\): a comprehensive examination of the metric\.Note:White Paper \#49Paper 1: §5 PUE comparison, §9 gaming risks\.Cited by:[§1\.6](https://arxiv.org/html/2605.22883#S1.SS6.p1.1),[§9\.5](https://arxiv.org/html/2605.22883#S9.SS5.p1.1)\.
- A\. Tschandet al\.\(2024\)MLPerf power: benchmarking the energy efficiency of machine learning systems fromμ\\muWatts to MWatts for sustainable AI\.External Links:2410\.12032Cited by:[§9\.1](https://arxiv.org/html/2605.22883#S9.SS1.p1.1),[Table 8](https://arxiv.org/html/2605.22883#S9.T8.65.60.1.6.1.1.1)\.
- A\. W\. van der Vaart \(1998\)Asymptotic statistics\.Cambridge Series in Statistical and Probabilistic Mathematics,Cambridge University Press\.External Links:[Document](https://dx.doi.org/10.1017/CBO9780511802256)Cited by:[Appendix F](https://arxiv.org/html/2605.22883#A6.SS0.SSS0.Px3.1.p1.1)\.
- White & Case LLP \(2025\)Energy efficiency requirements under the EU AI Act\.Technical reportWhite & Case LLP\.Cited by:[§1\.5](https://arxiv.org/html/2605.22883#S1.SS5.p2.1)\.

## Appendix AA\-LEMS System Architecture

Layer 1Layer 2Layer 3Layer 4RAPL100 Hzperf10 HzThermal1 HzNon\-blocking Queue\(oldest\-drop policy\)WorkloadInstrumentationSQLite \(53 tables\)energy\_samples, orch\_events,methodology\_registryAsync ETL Pipelineattribution \+ coverage \+ provenanceSQL Views \+ Registry API\(researcher\-facing\)Figure 14\.A\-LEMSfour\-layer architecture\. Layer 1: multi\-rate hardware collectors\. Layer 2: non\-blocking queue \+ workload instrumentation\. Layer 3: structured SQLite storage \(53 tables, analytical views\)\. Layer 4: async ETL \+ methodology registry\.A\-LEMSis implemented as a Python measurement harness running on the same machine as the workload under study\. The collector samplesRAPLenergy at 100 Hz via a non\-blocking queue with oldest\-sample\-drop policy \(prevents backpressure at high load\)\.energy\_samplesstores raw cumulativeRAPLstart/end values per interval, enabling post\-hoc recomputation with different attribution models\.

The ETL pipeline runs three async jobs after each experiment pair: phase attribution \(alignsenergy\_sampleswithorchestration\_events\), hardware metrics aggregation, and energy attribution \(computes all L0–L4 values\)\. The methodology registry stores every formula, confidence score, and literature reference in a machine\-readable database, queryable at runtime without source code access\.

## Appendix BExperimental Configuration Files

### B\.1\.Failure Injection Study Configuration

```
# config/experiment_configs/failure_injection_study.yaml
study:
  name: "Failure Injection Study"
  experiment_type: "failure_injection"
  experiment_goal: "Measure energy cost of failures and retry recovery"

tasks:
  - id: science_qa
  - id: tg_single_db
  - id: tg_single_calc
  - id: gsm8k_multi_step

execution:
  repetitions: 30
  cool_down_seconds: 30
  save_db: true

retry_policy:
  max_retries: 5
  retry_on_timeout: true
  retry_on_tool_error: true
  retry_on_api_error: true

failure_injection:
  enabled: true
  tool_failure_rate: 0.5
  timeout_rate: 0.5
```

Figure 15\.Failure injection configuration used in Section 8 experiments\.

## Appendix CResearch Query Library

The following canonical SQL queries are stored inconfig/research\_queries\.sqland are reproduced here for transparency\. Researchers can reproduce all paper figures and tables directly from theA\-LEMSSQLite database\.

Listing 1:RQ\-01: Core EpG query across workflow types\.1SELECT

2ge\.workflow\_type,

3AVG\(ge\.overhead\_fraction\)ASavg\_overhead\_fraction,

4AVG\(ge\.orchestration\_fraction\)ASavg\_orchestration\_fraction,

5AVG\(ge\.total\_energy\_uj\)/1e6ASavg\_epg\_j,

6COUNT\(\*\)ASn\_goals,

7SUM\(ge\.success\)\*1\.0/COUNT\(\*\)ASsuccess\_rate

8FROMgoal\_executionge

9JOINexperimentseONge\.exp\_id=e\.exp\_id

10WHEREe\.experiment\_type\!=’debug’

11ANDe\.experiment\_valid=1

12GROUPBYge\.workflow\_type;

Listing 2:RQ\-02: OOI per paired experiment\.1SELECT

2ge\_ag\.goal\_description,

3ge\_ag\.total\_energy\_uj/1e6ASepg\_agentic\_j,

4ge\_lin\.total\_energy\_uj/1e6ASepg\_linear\_j,

5ROUND\(ge\_ag\.total\_energy\_uj\*1\.0

6/NULLIF\(ge\_lin\.total\_energy\_uj,0\),3\)ASooi

7FROMgoal\_executionge\_ag

8JOINgoal\_executionge\_lin

9ONge\_ag\.task\_id=ge\_lin\.task\_id

10ANDge\_lin\.workflow\_type=’linear’

11WHEREge\_ag\.workflow\_type=’agentic’

12ANDge\_ag\.total\_energy\_ujISNOTNULL

13ANDge\_lin\.total\_energy\_ujISNOTNULL

14ORDERBYooiDESC;

Listing 3:RQ\-03: Phase power from RAPL sample alignment\.1SELECT

2oe\.phase,

3oe\.event\_type,

4ROUND\(\(oe\.end\_time\_ns\-oe\.start\_time\_ns\)/1e6,2\)ASms,

5COUNT\(es\.sample\_id\)ASsamples,

6ROUND\(AVG\(

7\(es\.pkg\_end\_uj\-es\.pkg\_start\_uj\)/

8CAST\(\(es\.sample\_end\_ns\-es\.sample\_start\_ns\)ASREAL\)

9\*1e3\),2\)ASavg\_power\_mw

10FROMorchestration\_eventsoe

11LEFTJOINenergy\_sampleses

12ONes\.run\_id=oe\.run\_id

13ANDes\.sample\_start\_ns\>=oe\.start\_time\_ns

14ANDes\.sample\_end\_ns<=oe\.end\_time\_ns

15WHEREoe\.run\_id=:run\_id

16GROUPBYoe\.event\_id

17ORDERBYoe\.start\_time\_ns;

Listing 4:RQ\-04: Boundary sensitivity — EpG under three boundary choices\.1SELECT

2ge\.workflow\_type,

3AVG\(r\.attributed\_energy\_uj\)/1e6ASstrict\_epg\_j,

4AVG\(r\.attributed\_energy\_uj

5\+COALESCE\(r\.pre\_task\_energy\_uj,0\)

6\)/1e6ASstandard\_epg\_j,

7AVG\(r\.attributed\_energy\_uj

8\+COALESCE\(r\.pre\_task\_energy\_uj,0\)

9\+COALESCE\(r\.post\_task\_energy\_uj,0\)

10\)/1e6ASloose\_epg\_j

11FROMgoal\_executionge

12JOINgoal\_attemptgaONga\.goal\_id=ge\.goal\_id

13JOINrunsrONr\.run\_id=ga\.run\_id

14JOINexperimentseONe\.exp\_id=ge\.exp\_id

15WHEREe\.is\_valid=1ANDe\.experiment\_type\!=’debug’

16ANDr\.attributed\_energy\_ujISNOTNULL

17GROUPBYge\.workflow\_type;

Listing 5:RQ\-05: Reproducibility — EpG variance per goal\.1SELECT

2ge\.goal\_description,

3COUNT\(\*\)ASn\_reps,

4ROUND\(AVG\(ge\.total\_energy\_uj\)/1e6,3\)ASmean\_epg\_j,

5ROUND\(

6100\.0\*\(MAX\(ge\.total\_energy\_uj\)

7\-MIN\(ge\.total\_energy\_uj\)\)

8/AVG\(ge\.total\_energy\_uj\),2\)ASrange\_pct

9FROMgoal\_executionge

10JOINexperimentseONge\.exp\_id=e\.exp\_id

11WHEREe\.experiment\_typeIN\(’normal’,’overhead\_study’\)

12ANDe\.is\_valid=1

13ANDge\.total\_energy\_ujISNOTNULL

14GROUPBYge\.task\_id

15HAVINGCOUNT\(\*\)\>=3

16ORDERBYrange\_pctASC;

Listing 6:RQ\-06: Measurement correctness — L1 validity\.1SELECT

2SUM\(CASEWHENr\.dynamic\_energy\_uj\>=0THEN1ELSE0END\)

3ASl1\_valid\_count,

4COUNT\(\*\)AStotal\_runs,

5ROUND\(

6100\.0\*SUM\(CASEWHENr\.dynamic\_energy\_uj\>=0

7THEN1ELSE0END\)

8/COUNT\(\*\),2\)ASvalidity\_pct,

9AVG\(r\.energy\_sample\_coverage\_pct\)ASavg\_coverage\_pct,

10SUM\(CASEWHENr\.energy\_sample\_coverage\_pct\>=95

11THEN1ELSE0END\)ASgold\_count,

12SUM\(CASEWHENr\.energy\_sample\_coverage\_pct\>=80

13ANDr\.energy\_sample\_coverage\_pct<95

14THEN1ELSE0END\)ASacceptable\_count

15FROMrunsr

16JOINexperimentseONr\.exp\_id=e\.exp\_id

17WHEREe\.is\_valid=1;

## Appendix DMethodology Registry Summary

Table[9](https://arxiv.org/html/2605.22883#A4.T9)summarises the full A\-LEMS measurement chain from raw hardware signals to derived metrics, with provenance tier and paper section cross\-references for each quantity\.

Table 9\.A\-LEMS measurement chain: derivation method, provenance tier, and paper usage for all primary quantities\. MEASURED = direct hardware read; CALCULATED = algebraic derivation; INFERRED = sample\-based estimation\.
## Appendix EPlatform Compliance Matrix

Table 10\.Platform support matrix forA\-LEMS\. ✓ = full;∼\\sim= partial;×\\times= not supported\.All timing usestime\.perf\_counter\(\): Python’s monotonic nanosecond\-resolution clock available on all platforms \(Table[10](https://arxiv.org/html/2605.22883#A5.T10)\); allRAPL\-derived metrics storeNULLwith graceful degradation\.

## Appendix FStochastic Workflow Model

#### Setup\.

We model a workflow unit as a stochastic process whose runtime structure is determined by per\-attempt success and per\-step energy realizations\. The model is platform\-agnostic: it does not assume any specific orchestration framework, provider, or task family\.

###### Definition 0 \(Workflow process\)\.

A workflow unit for goalggon system𝒮\\mathcal\{S\}is a tuple𝒲g=\(K,\{Xk\}k=1K,\{Ek\}k=1K\)\\mathcal\{W\}\_\{g\}=\(K,\\\{X\_\{k\}\\\}\_\{k=1\}^\{K\},\\\{E\_\{k\}\\\}\_\{k=1\}^\{K\}\)where:

- •K∈\{1,2,…,Kmax\}∪\{∞\}K\\in\\\{1,2,\\dots,K\_\{\\max\}\\\}\\cup\\\{\\infty\\\}is the \(random\) number of attempts before termination,
- •Xk∈\{0,1\}X\_\{k\}\\in\\\{0,1\\\}is the success indicator of attemptkk,
- •Ek∈ℝ≥0E\_\{k\}\\in\\mathbb\{R\}\_\{\\geq 0\}is the energy of attemptkk\(joules\),
- •Kmax∈ℕK\_\{\\max\}\\in\\mathbb\{N\}is a system\-imposed retry budget\.

We say the workflow*succeeds*iffXK=1X\_\{K\}=1andK≤KmaxK\\leq K\_\{\\max\}, and write𝟙g=1\\mathds\{1\}\_\{g\}=1; otherwise𝟙g=0\\mathds\{1\}\_\{g\}=0\.

The total energy of a workflow is:

Ewf,g=∑k=1KEk\.E\_\{\\mathrm\{wf\},g\}=\\sum\_\{k=1\}^\{K\}E\_\{k\}\.
We assume\{Ek\}\\\{E\_\{k\}\\\}are i\.i\.d\. across attempts with𝔼​\[Ek\]=μE\\mathbb\{E\}\[E\_\{k\}\]=\\mu\_\{E\}andVar​\(Ek\)=σE2\\mathrm\{Var\}\(E\_\{k\}\)=\\sigma\_\{E\}^\{2\}\. We assume per\-attempt success follows a Bernoulli process with parameterp∈\(0,1\]p\\in\(0,1\], i\.e\.,Xk∼Bern​\(p\)X\_\{k\}\\sim\\mathrm\{Bern\}\(p\)i\.i\.d\.

222Independence ofXkX\_\{k\}is a modeling assumption; retry behavior in real systems may induce dependence via prompt drift or state conditioning\. We analyze this empirically in §[8](https://arxiv.org/html/2605.22883#S8)Under this process,KKis a truncated geometric random variable:

Pr⁡\(K=k\)=\{\(1−p\)k−1​p,1≤k≤Kmax−1,\(1−p\)Kmax−1,k=Kmax\.\\Pr\(K=k\)=\\begin\{cases\}\(1\-p\)^\{k\-1\}p,&1\\leq k\\leq K\_\{\\max\}\-1,\\\\ \(1\-p\)^\{K\_\{\\max\}\-1\},&k=K\_\{\\max\}\.\\end\{cases\}
The success probability is:

π​\(p,Kmax\)=Pr⁡\(𝟙g=1\)=1−\(1−p\)Kmax\.\\pi\(p,K\_\{\\max\}\)=\\Pr\(\\mathds\{1\}\_\{g\}=1\)=1\-\(1\-p\)^\{K\_\{\\max\}\}\.

#### Closed\-form expectations\.

###### Theorem 2\(Expected workflow energy under truncated retry\)\.

Let𝒲g\\mathcal\{W\}\_\{g\}follow Definition[1](https://arxiv.org/html/2605.22883#A6.Thmtheorem1)\. Then:

𝔼​\[Ewf,g\]=μE⋅𝔼​\[K\]=μE⋅1−\(1−p\)Kmaxp\.\\mathbb\{E\}\[E\_\{\\mathrm\{wf\},g\}\]=\\mu\_\{E\}\\cdot\\mathbb\{E\}\[K\]=\\mu\_\{E\}\\cdot\\frac\{1\-\(1\-p\)^\{K\_\{\\max\}\}\}\{p\}\.

###### Proof\.

By Wald’s identity,𝔼​\[Ewf,g\]=μE​𝔼​\[K\]\\mathbb\{E\}\[E\_\{\\mathrm\{wf\},g\}\]=\\mu\_\{E\}\\mathbb\{E\}\[K\]\. Evaluating the truncated geometric expectation yields the result\. ∎

###### Corollary 0\(Untruncated limit\)\.

AsKmax→∞K\_\{\\max\}\\to\\infty,𝔼​\[K\]→1/p\\mathbb\{E\}\[K\]\\to 1/p\.

#### EpG as a ratio estimator\.

###### Definition 0 \(Population EpG\)\.

For populationΠ\\Piwith measureν\\nu:

EpG⋆=𝔼ν​\[Ewf,g\]𝔼ν​\[𝟙g\]=𝔼ν​\[μE⋅1−\(1−p\)Kmaxp\]𝔼ν​\[1−\(1−p\)Kmax\]\.\\mathrm\{EpG\}^\{\\star\}=\\frac\{\\mathbb\{E\}\_\{\\nu\}\[E\_\{\\mathrm\{wf\},g\}\]\}\{\\mathbb\{E\}\_\{\\nu\}\[\\mathds\{1\}\_\{g\}\]\}=\\frac\{\\mathbb\{E\}\_\{\\nu\}\\\!\\left\[\\mu\_\{E\}\\cdot\\frac\{1\-\(1\-p\)^\{K\_\{\\max\}\}\}\{p\}\\right\]\}\{\\mathbb\{E\}\_\{\\nu\}\\\!\\left\[1\-\(1\-p\)^\{K\_\{\\max\}\}\\right\]\}\.

Empirical estimator:

EpG^N=∑j=1NEwf,j∑j=1N𝟙j\.\\widehat\{\\mathrm\{EpG\}\}\_\{N\}=\\frac\{\\sum\_\{j=1\}^\{N\}E\_\{\\mathrm\{wf\},j\}\}\{\\sum\_\{j=1\}^\{N\}\\mathds\{1\}\_\{j\}\}\.
###### Proposition 0\(Consistency and asymptotic normality\)\.

AssumePr⁡\(𝟙g=1\)\>0\\Pr\(\\mathds\{1\}\_\{g\}=1\)\>0and𝔼​\[Ewf,g2\]<∞\\mathbb\{E\}\[E\_\{\\mathrm\{wf\},g\}^\{2\}\]<\\infty\. Then:

EpG^N→a\.s\.EpG⋆,N​\(EpG^N−EpG⋆\)→𝑑𝒩​\(0,VEpG\)\.\\widehat\{\\mathrm\{EpG\}\}\_\{N\}\\xrightarrow\{a\.s\.\}\\mathrm\{EpG\}^\{\\star\},\\quad\\sqrt\{N\}\(\\widehat\{\\mathrm\{EpG\}\}\_\{N\}\-\\mathrm\{EpG\}^\{\\star\}\)\\xrightarrow\{d\}\\mathcal\{N\}\(0,V\_\{\\mathrm\{EpG\}\}\)\.

###### Proof\.

Apply SLLN and delta method for ratio of means\(van der Vaart,[1998](https://arxiv.org/html/2605.22883#bib.bib39)\)\. ∎

#### Failure monotonicity and retry sensitivity\.

###### Proposition 0\(Failure monotonicity\)\.

HoldingμE,Kmax\\mu\_\{E\},K\_\{\\max\}fixed,EpG⋆\\mathrm\{EpG\}^\{\\star\}decreases inpp\.

###### Proposition 0\(Retry sensitivity\)\.

HoldingμE,p\\mu\_\{E\},pfixed,EpG⋆\\mathrm\{EpG\}^\{\\star\}is non\-decreasing inKmaxK\_\{\\max\}, with strict increase under non\-i\.i\.d\. attempts\.

#### OOI as paired estimator\.

###### Definition 0 \(Population OOI\)\.

OOI⋆=EpGA⋆EpGL⋆\.\\mathrm\{OOI\}^\{\\star\}=\\frac\{\\mathrm\{EpG\}\_\{A\}^\{\\star\}\}\{\\mathrm\{EpG\}\_\{L\}^\{\\star\}\}\.

###### Proposition 0\(OOI asymptotic distribution\)\.

N​\(OOI^N−OOI⋆\)→𝑑𝒩​\(0,VOOI\)\.\\sqrt\{N\}\(\\widehat\{\\mathrm\{OOI\}\}\_\{N\}\-\\mathrm\{OOI\}^\{\\star\}\)\\xrightarrow\{d\}\\mathcal\{N\}\(0,V\_\{\\mathrm\{OOI\}\}\)\.

#### Jensen lower bound\.

###### Proposition 0\(Convexity bound\)\.

EpG⋆≥𝔼​\[μE\]p¯\.\\mathrm\{EpG\}^\{\\star\}\\geq\\frac\{\\mathbb\{E\}\[\\mu\_\{E\}\]\}\{\\bar\{p\}\}\.

#### Summary\.

EpG is a ratio estimator over a stochastic workflow process with provable consistency, bounded expectations, and measurable sensitivity to system reliability structure \(Figure[16](https://arxiv.org/html/2605.22883#A6.F16)\)\.

![Refer to caption](https://arxiv.org/html/2605.22883v1/x8.png)Figure 16\.Empirical convergence ofEpG^N\\widehat\{\\mathrm\{EpG\}\}\_\{N\}asNNgrows\. Shaded bands show 95% bootstrap CIs at each subsample size; the1/N1/\\sqrt\{N\}contraction predicted by Proposition[5](https://arxiv.org/html/2605.22883#A6.Thmtheorem5)is visible\.

## Appendix GExecution Ontology

A\-LEMSdefines a four\-level execution hierarchy stored in the database:*Experiment*→\\to*Run*\(linear or agentic\)→\\to*Phase*\(planning/execution/synthesis\)→\\to*Event*\(LLM call, tool call, retry\)\. Each level maps to database tables with foreign key relationships, enabling energy attribution at any granularity\.

Thegoal\_executiontable is the paper’s fundamental unit of analysis: one row per user goal across one experiment\. A goal may require multiple attempts \(retries\)\. This table aggregates all attempt outcomes into a single success/failure verdict with full energy accounting\. ETL columns \(e\.g\.,total\_energy\_uj,overhead\_fraction\) are populated asynchronously after each experiment bygoal\_execution\_etl\.py\.

Energy conservation is enforced at the goal level:Eworkflow=∑iEattempt,iE\_\{\\mathrm\{workflow\}\}=\\sum\_\{i\}E\_\{\\mathrm\{attempt\},i\}, verified within 1 mJ tolerance for all goals with successful completions\. Any violation triggers aconservation\_violationflag inrun\_qualityand excludes the run from all research queries\.