Signal-Driven Observation for Long-Horizon Web Agents

# Signal-Driven Observation for Long-Horizon Web Agents
Source: [https://arxiv.org/html/2606.06708](https://arxiv.org/html/2606.06708)
###### Abstract

Web agents operating over long horizons ingest raw DOM and accessibility trees — routinely tens of thousands of tokens — at every action step, causing progressive context degradation that erodes reasoning well before tasks complete\. We argue that this coupling of observation frequency to action frequency is an architectural mistake\. Drawing on the insight from Recursive Language Models that querying a document outperforms reading it wholesale, we proposeSignal\-Driven Observation \(SDO\): a dedicated sub\-call reads the full DOM but returns only task\-relevant elements and their selectors, and is re\-invoked only when a lightweight signal detector fires — triggered by URL transitions, newly visible interactive elements, action failures, or exogenous browser events\. We outline the open problems SDO introduces and call on the community to treat observation compression as a core architectural decision in web agent design\.

web agents, observation compression, failure modes, long\-horizon tasks, context degradation, signal detection, recursive language models, DOM observation, trace diagnostics, agentic AI

## 1Introduction

Web agents are increasingly deployed on long\-horizon tasks—booking travel, completing multi\-page forms, navigating enterprise workflows—yet they fail at rates that make autonomous operation impractical\. In WebArena\(Zhouet al\.,[2023](https://arxiv.org/html/2606.06708#bib.bib1)\), even frontier agents achieve roughly 50% success, with the majority failures occurring mid\-task rather than at the outset\(Wanget al\.,[2026a](https://arxiv.org/html/2606.06708#bib.bib2)\)\. The pattern is consistent across benchmarks: agents get trapped in action loops, lose track of their original objective, and produce increasingly incoherent behavior as interaction history accumulates\(Chunget al\.,[2025](https://arxiv.org/html/2606.06708#bib.bib3)\)\.

The community has largely attributed these failures to context length limitations and responded with longer windows, summarization pipelines, and memory modules\. We argue that this diagnosis is incomplete\. The problem is not that agents run out of context\. The problem is that agents are*architecturally required to ingest everything they observe*, whether or not it is relevant to deciding the next action\.

Consider a concrete example\. A web agent at stepttreceives the raw accessibility tree of the current page—often 20,000 to 80,000 tokens—a screenshot, and its full interaction history\. It then invokes a complete LLM forward pass to decide a single action that may be as simple asfill\(\#email, "john@example\.com"\)\. The information required to make that decision is a single fact: the email field is focused and empty\. The remaining tens of thousands of tokens are not merely wasted—they actively degrade the model’s ability to reason about what matters\. This is not a context length problem\. It is an observation problem\.

We call this failure mode*observation over\-ingestion*: the architectural coupling of observation frequency to action frequency, forcing the agent to re\-read the full page state at every step regardless of whether that state has meaningfully changed\. Observation over\-ingestion is a triggering precondition for three well\-documented downstream failures\. First,*context rot*—the progressive degradation of reasoning quality as irrelevant observation tokens accumulate in the context window\. Second,*loop\-trapping*—the agent repeats the same action sequence because its bloated context prevents it from recognizing it has visited this state before\. Third,*goal drift*—the agent’s original objective is buried under layers of accumulated DOM noise, and it begins pursuing a different, emergent sub\-goal\(Arikeet al\.,[2025](https://arxiv.org/html/2606.06708#bib.bib4); Menonet al\.,[2026](https://arxiv.org/html/2606.06708#bib.bib5)\)\.

This failure mode is reproducible and independent of the model capability\. It occurs in frontier models with 200K\-token windows just as it occurs in smaller models, because the issue is not capacity, but architecture\(Yeet al\.,[2025](https://arxiv.org/html/2606.06708#bib.bib7)\)\. It is also invisible to standard evaluation: terminal success metrics record that the agent failed but not that observation over\-ingestion was the mechanism\. Trace\-level diagnostics that could surface this failure—logging when the observation changed versus when the agent re\-read it unchanged—do not exist in current evaluation frameworks\.

Recursive Language Models \(RLMs\)\(Zhanget al\.,[2025a](https://arxiv.org/html/2606.06708#bib.bib6)\)offer an architectural insight that has not yet been applied to this problem\. RLMs demonstrated that for large static documents, models perform dramatically better when they query a document programmatically rather than ingesting it wholesale\. The model treats the document as an external variable, calling sub\-processes to extract exactly what it needs\. The live DOM of a web page is precisely the environment where this insight is most needed—and most absent\.

We sketch one concrete instantiation of this principle—*Signal\-Driven Observation*\(SDO\)—not as a complete system but to demonstrate that observation over\-ingestion is architecturally avoidable and to surface the open design questions a solution must address\. In SDO, a dedicated sub\-call reads the full DOM but returns only a compact, task\-conditioned summary\. This sub\-call is not invoked at every action step\. A lightweight signal detector monitors four browser\-native events—URL transitions, newly visible interactive elements, action failures, and exogenous page events—and triggers re\-observation only when the page state has meaningfully changed\. Between signals, the root model executes its planned action sequence with no additional LLM calls and no growing context\.

This paper makes three contributions\. First, we define observation over\-ingestion as an operational failure mode in web agents—a reproducible triggering precondition for context rot, loop\-trapping, and goal drift that is distinct from context length exhaustion\. Second, we sketch Signal\-Driven Observation as a concrete demonstration that observation frequency can be decoupled from action frequency\. Third, we identify the open problems this framing surfaces and call on the community to treat observation compression as a first\-class target for failure mitigation in agentic systems\.

## 2Related Work

### 2\.1Web Agent Benchmarks and Long\-Horizon Evaluation

The evaluation landscape for web agents has expanded rapidly\. WebArena\(Zhouet al\.,[2023](https://arxiv.org/html/2606.06708#bib.bib1)\)introduced self\-hosted, reproducible web environments with programmatic evaluation; WorkArena\(Drouinet al\.,[2024](https://arxiv.org/html/2606.06708#bib.bib8)\), WorkArena\+\+\(Boisvertet al\.,[2024](https://arxiv.org/html/2606.06708#bib.bib9)\)extended this to enterprise ServiceNow workflows where single HTML pages can reach 40K–500K tokens\. BrowserGym\(de Chezelleset al\.,[2024](https://arxiv.org/html/2606.06708#bib.bib10)\)provides the unified observation framework underlying several of these benchmarks, exposing the full accessibility tree, raw HTML, and viewport screenshots at every step—a single observation routinely exceeding 20,000 tokens\. VisualWebArena\(Kohet al\.,[2024](https://arxiv.org/html/2606.06708#bib.bib11)\)and OSWorld\(Xieet al\.,[2024](https://arxiv.org/html/2606.06708#bib.bib12)\)broadened evaluation to multimodal and desktop\-level tasks respectively\. More recently, Online\-Mind2Web\(Xueet al\.,[2025](https://arxiv.org/html/2606.06708#bib.bib13)\)demonstrated that frontier agents are up to 59% less competent on live websites than static benchmarks suggest, attributing much of the gap to dynamic content and evaluation artifacts\. REAL\(Garget al\.,[2025](https://arxiv.org/html/2606.06708#bib.bib14)\)provides deterministic website replicas for reproducible evaluation, while Odysseys\(Janget al\.,[2026](https://arxiv.org/html/2606.06708#bib.bib15)\)introduces 200 long\-horizon, multi\-site tasks derived from real browsing histories and explicitly identifies the inadequacy of trajectory\-level LLM\-as\-judge evaluation for long tasks\.

The pattern across these benchmarks is consistent: as task horizon grows, agent success degrades sharply\. HORIZON\(Wanget al\.,[2026b](https://arxiv.org/html/2606.06708#bib.bib16)\)formalizes this with a cross\-domain diagnostic benchmark and trajectory\-grounded failure attribution, showing that long\-horizon failures are systematically under\-characterized by terminal metrics\.\(Chunget al\.,[2025](https://arxiv.org/html/2606.06708#bib.bib3)\)find that agents primarily fail by getting stuck in loops and losing track of objectives—not from exhausting context tokens\. A subgoal\-driven analysis\(Wanget al\.,[2026a](https://arxiv.org/html/2606.06708#bib.bib2)\)reports mid\-task stuck behavior in nearly 50% of WebArena\-Lite trajectories for Gemini\-2\.5\-Pro\. These findings motivate our claim that the failure mechanism lies in*how*agents observe, not*how much*they can store\.

### 2\.2Context and Observation Management

A growing body of work addresses the observation burden directly\. AgentFold\(Yeet al\.,[2025](https://arxiv.org/html/2606.06708#bib.bib7)\)reports that approximately 20% of long\-horizon tasks are forcibly terminated at 100 turns despite contexts using only around 7K tokens, suggesting failures arise from in\-context confusion rather than raw length\. FocusAgent\(Kerbouaet al\.,[2025a](https://arxiv.org/html/2606.06708#bib.bib17)\)uses a lightweight LLM retriever to extract task\-relevant lines from accessibility\-tree observations, reducing observation size by over 50% on WorkArena and WebArena while also lowering prompt\-injection success rates\. LineRetriever\(Kerbouaet al\.,[2025b](https://arxiv.org/html/2606.06708#bib.bib18)\)proposes planning\-aware observation reduction via embedding\-based retrieval, explicitly motivated by the insight that agents “need future\-action\-relevant context, not just semantically similar text\.” ACON\(Kanget al\.,[2025](https://arxiv.org/html/2606.06708#bib.bib19)\)provides a unified framework that compresses both observations and interaction histories through failure\-driven natural\-language guidelines, achieving 26–54% peak\-token reduction on AppWorld\(Trivediet al\.,[2024](https://arxiv.org/html/2606.06708#bib.bib20)\)and OfficeBench\(Wanget al\.,[2024](https://arxiv.org/html/2606.06708#bib.bib21)\)\. Hierarchical Memory Tree\(Tanet al\.,[2026](https://arxiv.org/html/2606.06708#bib.bib22)\)abstracts raw HTML trajectories into compact semantic descriptions, reporting 72\.7% context\-length reduction on WebArena\. SLIM\(Yenet al\.,[2025](https://arxiv.org/html/2606.06708#bib.bib23)\)introduces summarization for accumulated search content, dual\-memory frameworks such as M2\(Yanet al\.,[2026](https://arxiv.org/html/2606.06708#bib.bib24)\)separate working from long\-term memory, and ContextBudget\(Wuet al\.,[2026](https://arxiv.org/html/2606.06708#bib.bib25)\)introduces dynamic budget\-conditioned compression that explicitly identifies failure modes of budget\-free approaches\.

A counterpoint is offered by\(Enomotoet al\.,[2026](https://arxiv.org/html/2606.06708#bib.bib26)\), who argue that the optimal observation representation depends on model capability—compact accessibility trees for weaker models, raw HTML with thinking budget for stronger ones\. This highlights that compression is not uniformly beneficial, a tension our work acknowledges\.

All of these approaches share a common assumption: the agent*should*observe the full page state and the engineering problem is compressing it afterward\. None questions whether full observation should occur at every step in the first place\. Observation over\-ingestion is treated as a data\-volume problem rather than a frequency problem\.

### 2\.3Failure Diagnosis, Safety, and Trace\-Level Attribution

The failure\-diagnosis literature has matured rapidly\. On the robustness side, WAREX\(Karaet al\.,[2025](https://arxiv.org/html/2606.06708#bib.bib27)\)injects network errors, server failures, and malicious popups into WebArena and REAL, demonstrating significant task\-success drops under realistic perturbations\. StressWeb\(Baiet al\.,[2026](https://arxiv.org/html/2606.06708#bib.bib28)\)introduces controlled perturbations across the perception, semantic, and execution stages of the interaction pipeline\. DoomArena\(Boisvertet al\.,[2025](https://arxiv.org/html/2606.06708#bib.bib29)\)attacks BrowserGym agents with popup injections hidden in accessibility attributes, achieving 23–78% failure rates on OSWorld\.\(Zhanget al\.,[2024](https://arxiv.org/html/2606.06708#bib.bib30)\)show that adversarial popups alone reach 86% click\-through rates and reduce task success by 47%\.

The safety implications of accumulated context are also emerging: MT\-AgentRisk\(Liet al\.,[2026](https://arxiv.org/html/2606.06708#bib.bib31)\)shows that multi\-turn observation histories create attack surfaces beyond the capability failures we focus on here\.

On the attribution side, AgentRx\(Schmidgall and Moor,[2025](https://arxiv.org/html/2606.06708#bib.bib32)\)provides a nine\-category failure taxonomy over 115 annotated failed trajectories\. AgenTracer\(Zhanget al\.,[2025b](https://arxiv.org/html/2606.06708#bib.bib33)\)uses counterfactual replay and programmed fault injection\. ST\-WebAgentBench\(Levyet al\.,[2024](https://arxiv.org/html/2606.06708#bib.bib34)\)introduces Completion\-under\-Policy, showing that safety\-adjusted success can be less than two\-thirds of nominal completion rates\.\(Fanget al\.,[2025](https://arxiv.org/html/2606.06708#bib.bib35)\)study preemptive detection and correction of misaligned actions in web shopping agents—exactly the kind of grounding error that bloated observations make harder to catch\.

These frameworks diagnose failures at the*action*level: which action was wrong, which policy was violated, which step was the root cause\. None diagnoses failures at the*observation*level—whether the agent re\-ingested an unchanged page, whether a fresh observation would have prevented the error, or whether the volume of observation tokens was itself the degradation mechanism\. If observation over\-ingestion is the triggering precondition, then action\-level attribution will consistently mislocate the cause\.

### 2\.4Recursive and Hierarchical Agent Architectures

Recursive Language Models \(RLMs\)\(Zhanget al\.,[2025a](https://arxiv.org/html/2606.06708#bib.bib6)\)treat the input prompt as an external variable in a REPL environment, recursively calling sub\-instances to peek at relevant slices without bloating the root context\. On long\-context benchmarks, RLMs maintain strong performance at inputs exceeding ten million tokens where standard models degrade sharply\. However, RLMs were designed for static documents—codebases, legal corpora, long\-form text\. The live DOM of a web page mutates with every agent action and can be interrupted by exogenous events independent of the agent’s behavior\. Applying recursive decomposition to a dynamic, stateful environment requires addressing when to re\-observe and what to extract, neither of which arises in the static setting\.

Prior web agent architectures have explored related structural ideas without addressing observation frequency\. LASER\(Maet al\.,[2023](https://arxiv.org/html/2606.06708#bib.bib36)\)introduces URL\-based state tracking with backtracking on e\-commerce tasks\. BEAP\-Agent\(Luet al\.,[2026](https://arxiv.org/html/2606.06708#bib.bib37)\)uses depth\-first search with multi\-level backtracking\. RCI\(Kimet al\.,[2023](https://arxiv.org/html/2606.06708#bib.bib38)\)adds self\-critique loops that evaluate action quality\. RAGEN\(Wanget al\.,[2025](https://arxiv.org/html/2606.06708#bib.bib39)\)introduces StarPO for training agentic reasoning and documents the “Echo Trap” failure mode—agents echoing earlier thoughts—which is an observation\-context pathology directly caused by accumulated context noise\. Subgoal\-driven frameworks\(Wanget al\.,[2026a](https://arxiv.org/html/2606.06708#bib.bib2)\)decompose long\-horizon tasks into explicit sub\-plans\. All of these operate*over*the observation—planning what to do with what the agent sees—but none intervenes on the observation itself\. At every action step, the full DOM is still read, the full context still grows, and the same over\-ingestion mechanism still applies\.

The gap is precise: no existing architecture decouples observation frequency from action frequency in web agents\. We propose Signal\-Driven Observation as one concrete way to close it\.

## 3Signal\-Driven Observation

We sketch*Signal\-Driven Observation*\(SDO\) \(Figure[1](https://arxiv.org/html/2606.06708#S3.F1)\)—not as a complete system but as a concrete instantiation of a principle: observation frequency should be decoupled from action frequency\. SDO makes this precise by defining a set of lightweight, zero\-LLM\-cost signals that determine when a new observation is warranted, and delegating all DOM reading to a focused sub\-call that the root LM never performs itself\. We frame each architectural component in terms of the failure mode it is designed to prevent\.

![Refer to caption](https://arxiv.org/html/2606.06708v1/figures/SDO_architecture.png)Figure 1:SDO architecture\. The Signal Detector runs after every action at zero LLM cost\. sub\_RLM is invoked only when a signal fires, returning a compact observationOt\+1O\_\{t\+1\}\. The Root LM replans from bounded context\.### 3\.1Architecture

SDO involves four components operating at runtime over a standard browser controlled via Playwright\.

#### Root LM\.

The root LM maintains three variables throughout the task: the original task specificationTT, which never changes; a sequence of compact observationsO1,O2,…O\_\{1\},O\_\{2\},\\ldotsreceived from sub\_RLM; and an action historyH=\[a1,a2,…\]H=\[a\_\{1\},a\_\{2\},\\ldots\]that grows by one entry after every executed action\. The root LM plans a sequence of actionsSSfrom\(T,Ot,H\)\(T,O\_\{t\},H\)and replans only when a new observation arrives\.

The failure\-mode logic is direct\. Because the root LM never reads raw DOM, its context contains only task\-relevant information: compact observations and single\-line action entries\. The signal\-to\-noise ratio in the root LM’s context remains high throughout the task, removing the primary triggering precondition for context rot\. And becauseTTis never buried under thousands of irrelevant DOM tokens, the root LM’s access to the original task objective is preserved, reducing the conditions under which goal drift emerges\.

#### sub\_RLM\.

sub\_RLM is an LM call—either the same model as the root LM under a different prompt, or a smaller dedicated model—whose sole responsibility is DOM observation\. Given the current page and the taskTT, it reads the full accessibility tree and returns a compact, task\-conditioned observationOtO\_\{t\}: a structured list of only the task\-relevant interactive elements, each described by its purpose, type, current value, and Playwright selector\. sub\_RLM does not plan or execute actions\. It answers one question:*what on this page matters for the current task, and where is it?*We describe the text\-only case for concreteness, but sub\_RLM can equally operate over a screenshot, or both modalities together—the architectural contribution is the decoupling of observation from action, not the specific modality of the sub\-call\.

The task\-conditioning is critical for failure prevention\. A raw accessibility tree contains every element on the page—navigation menus, footer links, sidebar widgets, hidden elements, advertising containers\. Each of these is noise that competes with the task\-relevant signal\. By filtering to only elements that matter forTT, sub\_RLM removes this noise at the source rather than asking the root LM to filter it under growing context pressure\.

#### Signal Detector\.

The signal detector runs after every Playwright action at zero LLM cost\. It monitors four browser\-native conditions:

1. 1\.URL transition—the page navigated to a new address, indicating a state change that invalidates the current observation\.
2. 2\.New ARIA element—an element with a role such aslistbox,select,menu,combobox,dialog, or witharia\-expanded=truebecame visible that was not present inOtO\_\{t\}\. This covers dropdowns opening, modals appearing, and dynamic widgets loading—all cases where new interactive content that the root LM needs to act on has appeared in the DOM\.
3. 3\.Action failure—the Playwright action raised an exception, timed out, or the targeted element became unavailable, indicating the page is not in the state the root LM assumed\.
4. 4\.Exogenous event—a DOM mutation occurred that was not caused by the agent’s action: a cookie consent banner injected by a JavaScript timer, an AJAX response rewriting page content, a session timeout redirect, or a third\-party script modifying layout\. Signals 1–3 arise in both simulated environments and live browsers\. Signal 4 arises primarily on real websites—a gap we discuss in Section[4](https://arxiv.org/html/2606.06708#S4)\.

Beyond its role in triggering re\-observation, the signal detector produces a structured trace of environment state changes across the trajectory\. This trace—which signals fired at which steps, and of which type—is itself a diagnostic artifact that current evaluation frameworks do not capture\. It enables a form of observation\-level attribution: when the agent fails at stepkk, the trace can reveal whether a signal was missed at an earlier step, whether the observation was stale, or whether a re\-observation was triggered but produced an insufficientOtO\_\{t\}\.

#### Browser\.

The browser executes Playwright actions from the sequenceSSproduced by the root LM\. After each action, control passes to the signal detector before the next action inSSbegins\. The root LM never interacts with the browser directly\.

### 3\.2The SDO Loop

On task start, the initial page load triggers a sub\_RLM call, producing the first observationO1O\_\{1\}before the root LM has taken any action\. The root LM receives\(T,O1,H=\[\]\)\(T,O\_\{1\},H\{=\}\[\]\)and plans an initial action sequenceSS\. For each actiona∈Sa\\in S: the browser executesaa; the action is appended toHH; the signal detector checks the four conditions\. If no signal fires, the next action inSSexecutes immediately—no LLM call, no DOM read\. If any signal fires, sub\_RLM is called on the current DOM, producingOt\+1O\_\{t\+1\}; the root LM receives the updated\(T,Ot\+1,H\)\(T,O\_\{t\+1\},H\)and replans from the new observation\.

### 3\.3Case Study: E\-Commerce Purchase

We trace a concrete long\-horizon task—*“Find a pair of wireless noise\-cancelling headphones under $200 on an e\-commerce site and complete checkout”*—to illustrate both how observation over\-ingestion causes cascading failures under the standard approach and how SDO prevents them\.

#### Standard agent \(without SDO\)\.

The agent lands on the homepage\. The raw DOM contains approximately 1,200 elements: navigation bars, promotional carousels, trending product grids, footer links, search bars, and login prompts\. The agent ingests all of it and decides to search for “wireless noise\-cancelling headphones\.”

*Step 1–3: Search and filter\.*The agent types the query and submits\. The results page contains 48 product listings, sponsored results, filter panels, pagination controls, and a sorting dropdown\. Each product listing includes title, price, rating, review count, and availability\. The agent ingests all of it, identifies a promising pair at $179, and clicks through\.

*Step 4–6: Product detail page\.*The product page includes specifications, customer reviews, related products, “frequently bought together” suggestions, seller information, and shipping options\. By this point the root LM’s context contains the homepage DOM, the results page DOM, and the product page DOM—three full raw snapshots, each tens of thousands of tokens, accumulating into a context dominated by irrelevant page content\. A cookie consent banner fires \(exogenous event\)\. The banner overlays the “Add to Cart” button\. The agent, whose context is already degraded, clicks the banner’s “Accept” button but interprets the resulting page state incorrectly—it believes it has added the item to cart\.

*Step 7–10: Cart and checkout\.*The agent navigates to the cart\. The cart is empty—the item was never added\. But the agent’s context now spans four full DOM snapshots, each tens of thousands of tokens\. The original task specification—*wireless noise\-cancelling headphones under $200*—is buried deep in the context\. The agent sees “Your cart is empty” but, rather than recognizing the failure and backtracking, it searches for a new product \(goal drift\)\. It finds a $350 pair of wired headphones— violating both the wireless and the price constraints\. By step 12, it has begun the checkout process for the wrong product\. Terminal evaluation records:*task failed\.*It does not record where or why\.

#### SDO agent\.

The agent lands on the same homepage\. sub\_RLM reads the 1,200 elements and returnsO1O\_\{1\}: a single task\-relevant element—the search bar, with its selector and purpose\. The root LM’s context contains the task and one compact observation\. It plans: type query, click search\.

*Step 1–3: Search and filter\.*After submitting the search, the URL changes—signal fires\. sub\_RLM reads the results page\. 48 listings exist, but only 6 are wireless noise\-cancelling headphones under $200\.O2O\_\{2\}contains those 6, each with name, price, and selector—a compact representation orders of magnitude smaller than the raw results page\. It selects the $179 pair and clicks\.

*Step 4–6: Product detail page\.*URL change—signal fires\. sub\_RLM returnsO3O\_\{3\}: product name, price \($179\), “Add to Cart” button with selector, and one unexpected element—a cookie consent banner flagged astask\_relevant: false, blocking: true\. Root LM sees this is an interruption, not a state change\. It plans: dismiss banner, then click Add to Cart\. Two actions, no LLM call between them\. After “Add to Cart,” a small AJAX confirmation appears \(new ARIA element—signal fires\)\. sub\_RLM confirms: “Item added\. Cart count: 1\.” The root LM’s context has grown only by the addition of compact observations and action entries—not by accumulated raw DOM\.

*Step 7–10: Cart and checkout\.*The agent navigates to cart\. URL change—signal fires\.O5O\_\{5\}: cart contains “Sony WH\-1000XM5, $179, quantity 1\.” Root LM compares againstTT: wireless noise\-cancelling, under $200—match\. Proceeds to checkout\. Each checkout page triggers one sub\_RLM call, returning only the form fields needed\. The agent completes checkout in 4 actions across 2 pages\.

*Outcome:*12 actions, 6 sub\_RLM calls\. The root LM’s context grew only at signal boundaries—each time by a compact observation of a few task\-relevant elements—rather than by tens of thousands of raw DOM tokens at every step\. The task specificationTTremained immediately accessible throughout the trajectory, never buried under irrelevant page content\. No context rot\. No goal drift\. The cookie banner was handled as an interruption, not a confusion point\.

#### Failure\-mode comparison\.

The standard agent failed not because any individual action was wrong, but because accumulated observation noise degraded its ability to maintain task coherence\. The cookie banner was the proximate trigger, but the precondition was multiple full DOM snapshots of irrelevant content already in context\. SDO prevents the precondition: the root LM never sees irrelevant DOM, so when the cookie banner appears, it arrives in a clean context where the root LM can correctly classify it as an interruption and plan around it\.

This illustrates the core claim: observation over\-ingestion is not merely inefficient—it is a failure mechanism\. Removing it does not just save tokens\. It preserves the conditions under which the root LM can reason correctly about the task, even when the environment introduces unexpected complications\.

## 4Open Problems

SDO is a sketch, not a solution\. It demonstrates that observation over\-ingestion is architecturally avoidable, but it surfaces a set of design questions that must be resolved before the principle can be instantiated reliably\. We organize these as open problems for the community, each with an explicit description of the tradeoff involved\.

#### Signal completeness versus cost\.

The four signals we define—URL transition, new ARIA element, action failure, and exogenous event—are lightweight and cover the most common state changes in browser environments\. But they are not exhaustive\. A page can change meaningfully without triggering any of these: a price updating silently via JavaScript, a product going out of stock with no visible DOM change, a form validation message appearing inside an existing element rather than as a new one\. Missing such changes means the root LM plans against a stale observation\. Adding more signals increases coverage but also increases the frequency of sub\_RLM calls, eroding the efficiency gain that motivates SDO in the first place\. The tradeoff is between observation freshness and call frequency, and the right balance likely depends on the task domain and the cost of acting on stale information\.

#### Task\-conditioned filtering fidelity\.

sub\_RLM must decide which elements are task\-relevant without knowing what the root LM will need\. If it filters too aggressively, it omits elements the root LM would have used—a form of information loss that is unrecoverable without a full re\-observation\. If it filters too conservatively, observations grow large and the compression benefit diminishes\. In structured environments like WorkArena, where task parameters map directly to form fields, the filtering problem is tractable\. On open\-ended real\-web tasks—*“find a good anniversary gift under $100”*—task relevance is ambiguous and sub\_RLM must make judgment calls that may not match the root LM’s reasoning\. This is a compression\-completeness tradeoff: how much information can be discarded at the observation layer without degrading the planning layer?

#### Semantic errors within a state\.

SDO’s signal detection catches structural and technical failures but not semantic ones\. If the root LM plans the wrong action— selecting “Software” when the task requires “Hardware”—and the action executes successfully with no structural change to the page, no signal fires\. The agent continues on the wrong path without any observation\-level indication that something went wrong\. This error is only caught downstream: either when a later sub\_RLM call reveals an inconsistency, or at terminal evaluation when the task is already failed\. A natural extension is a lightweight value check after form\-filling actions—comparing the executed value against the planned value via a direct DOM read—but this check is reliable only in environments with stable selectors and standard HTML form elements\. On real websites using React\-controlled inputs or custom components, the DOM value and the application state can diverge\. Detecting mid\-sequence semantic drift without structural signals remains an open problem\.

#### Simulation versus real\-web gap\.

Signals 1–3 \(URL transition, new ARIA element, action failure\) arise naturally in both simulated benchmarks and live browsers\. Signal 4 \(exogenous events\) exposes a fundamental gap: simulated environments such as WebArena and WorkArena are deterministic by design—no cookie banners, no AJAX surprises, no session timeouts\. SDO’s signal detector can capture exogenous events on live browsers using standard Playwright event listeners and DOM mutation observers, requiring no model training\. However, this gap has broader implications for the community: agents that are trained via reinforcement learning in deterministic simulation never encounter exogenous events during training and therefore have no learned recovery behavior when deployed on real websites\. SDO sidesteps this for inference—the signal detector is rule\-based, not learned—but whether the root LM can reliably plan around exogenous interruptions it has never seen in its pretraining or prompting context remains an open question\.

#### Observation\-level evaluation\.

Current web agent benchmarks evaluate at the action level \(was the action correct?\) or the task level \(did the agent succeed?\)\. SDO motivates a third level:*observation\-level evaluation*\. Was the observation fresh when the agent needed it? Did the agent act on stale information? Would a re\-observation at stepkkhave prevented the failure at stepk\+3k\{\+\}3? The signal detector’s trace provides the raw data for this kind of analysis, but the metrics and benchmarks that would operationalize it do not yet exist\. Building them is a prerequisite for measuring whether observation compression actually prevents the failures it is designed to address, rather than merely reducing token count\.

#### Observation history management\.

As the trajectory lengthens, the list of past observationsO1,O2,…,OkO\_\{1\},O\_\{2\},\\ldots,O\_\{k\}grows in the root LM’s context\. While each individual observation is compact, the cumulative history can still become substantial over very long horizons—tasks spanning dozens of pages and hundreds of actions\. At some point, older observations may need to be summarized or discarded, reintroducing a version of the context management problem SDO was designed to avoid\. How to manage observation history without losing information needed for backtracking or loop detection is an open design question\.

## 5Conclusion

We have argued that observation over\-ingestion—the architectural coupling of observation frequency to action frequency—is a first\-class failure mode in web agents, distinct from context length exhaustion and serving as a reproducible triggering precondition for context rot, loop\-trapping, and goal drift\. We sketched Signal\-Driven Observation as one concrete demonstration that this coupling is not inevitable: a lightweight signal detector can govern when re\-observation occurs, and a focused sub\-call can compress what the agent sees without the root model ever reading raw DOM\. The approach is a direction, not a finished system, and we have been explicit about what it cannot yet handle—semantic errors without structural signals, filtering fidelity on open\-ended tasks, and the absence of exogenous events from simulated training environments\. We also note that this paper does not include experiments\. Whether SDO prevents the failures we describe in practice is a question that can only be answered by building and evaluating it—a natural next step that we hope this framing motivates\.

The broader point is simple\. The community has invested heavily in making agents reason better over what they observe\. Comparatively little attention has gone to whether agents should be observing what they observe in the first place\. We believe observation compression deserves to be treated as a first\-class design primitive—not a post\-hoc optimization, but a structural decision made before the first action is taken\. We hope this paper motivates others to formalize the problem and build the evaluation infrastructure needed to measure whether solving it actually prevents the failures it is designed to address\.

## References

- R\. Arike, E\. Donoway, H\. Bartsch, and M\. Hobbhahn \(2025\)Technical report: evaluating goal drift in language model agents\.CoRRabs/2505\.02709\.External Links:[Link](https://doi.org/10.48550/arXiv.2505.02709),[Document](https://dx.doi.org/10.48550/ARXIV.2505.02709),2505\.02709Cited by:[§1](https://arxiv.org/html/2606.06708#S1.p4.1)\.
- H\. Bai, D\. Wang, L\. Chen, B\. Hao, P\. Shao, Y\. Yang, Y\. He, and C\. Zhuang \(2026\)StressWeb: a diagnostic benchmark for web agent robustness under realistic interaction variability\.External Links:2604\.16385,[Link](https://arxiv.org/abs/2604.16385)Cited by:[§2\.3](https://arxiv.org/html/2606.06708#S2.SS3.p1.1)\.
- L\. Boisvert, M\. Bansal, C\. K\. R\. Evuru, G\. Huang, A\. Puri, A\. Bose, M\. Fazel, Q\. Cappart, J\. Stanley, A\. Lacoste, A\. Drouin, and K\. Dvijotham \(2025\)DoomArena: A framework for testing AI agents against evolving security threats\.CoRRabs/2504\.14064\.External Links:[Link](https://doi.org/10.48550/arXiv.2504.14064),[Document](https://dx.doi.org/10.48550/ARXIV.2504.14064),2504\.14064Cited by:[§2\.3](https://arxiv.org/html/2606.06708#S2.SS3.p1.1)\.
- L\. Boisvert, M\. Thakkar, M\. Gasse, M\. Caccia, T\. L\. S\. de Chezelles, Q\. Cappart, N\. Chapados, A\. Lacoste, and A\. Drouin \(2024\)WorkArena\+\+: towards compositional planning and reasoning\-based common knowledge work tasks\.CoRRabs/2407\.05291\.External Links:[Link](https://doi.org/10.48550/arXiv.2407.05291),[Document](https://dx.doi.org/10.48550/ARXIV.2407.05291),2407\.05291Cited by:[§2\.1](https://arxiv.org/html/2606.06708#S2.SS1.p1.1)\.
- A\. Chung, Y\. Zhang, K\. Lin, A\. Rawal, Q\. Gao, and J\. Chai \(2025\)Evaluating long\-context reasoning in llm\-based webagents\.CoRRabs/2512\.04307\.External Links:[Link](https://doi.org/10.48550/arXiv.2512.04307),[Document](https://dx.doi.org/10.48550/ARXIV.2512.04307),2512\.04307Cited by:[§1](https://arxiv.org/html/2606.06708#S1.p1.1),[§2\.1](https://arxiv.org/html/2606.06708#S2.SS1.p2.1)\.
- T\. L\. S\. de Chezelles, M\. Gasse, A\. Drouin, M\. Caccia, L\. Boisvert, M\. Thakkar, T\. Marty, R\. Assouel, S\. O\. Shayegan, L\. K\. Jang, X\. H\. Lù, O\. Yoran, D\. Kong, F\. F\. Xu, S\. Reddy, Q\. Cappart, G\. Neubig, R\. Salakhutdinov, N\. Chapados, and A\. Lacoste \(2024\)The browsergym ecosystem for web agent research\.CoRRabs/2412\.05467\.External Links:[Link](https://doi.org/10.48550/arXiv.2412.05467),[Document](https://dx.doi.org/10.48550/ARXIV.2412.05467),2412\.05467Cited by:[§2\.1](https://arxiv.org/html/2606.06708#S2.SS1.p1.1)\.
- A\. Drouin, M\. Gasse, M\. Caccia, I\. H\. Laradji, M\. D\. Verme, T\. Marty, L\. Boisvert, M\. Thakkar, Q\. Cappart, D\. Vázquez, N\. Chapados, and A\. Lacoste \(2024\)WorkArena: how capable are web agents at solving common knowledge work tasks?\.CoRRabs/2403\.07718\.External Links:[Link](https://doi.org/10.48550/arXiv.2403.07718),[Document](https://dx.doi.org/10.48550/ARXIV.2403.07718),2403\.07718Cited by:[§2\.1](https://arxiv.org/html/2606.06708#S2.SS1.p1.1)\.
- M\. Enomoto, R\. Obara, H\. Zhang, and M\. Oyamada \(2026\)Read more, think more: revisiting observation reduction for web agents\.CoRRabs/2604\.01535\.External Links:[Link](https://doi.org/10.48550/arXiv.2604.01535),[Document](https://dx.doi.org/10.48550/ARXIV.2604.01535),2604\.01535Cited by:[§2\.2](https://arxiv.org/html/2606.06708#S2.SS2.p2.1)\.
- H\. Fang, X\. Zhu, and I\. Gurevych \(2025\)Preemptive detection and correction of misaligned actions in LLM agents\.InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, EMNLP 2025, Suzhou, China, November 4\-9, 2025,C\. Christodoulopoulos, T\. Chakraborty, C\. Rose, and V\. Peng \(Eds\.\),pp\. 222–244\.External Links:[Link](https://doi.org/10.18653/v1/2025.emnlp-main.12),[Document](https://dx.doi.org/10.18653/V1/2025.EMNLP-MAIN.12)Cited by:[§2\.3](https://arxiv.org/html/2606.06708#S2.SS3.p3.1)\.
- D\. Garg, S\. VanWeelden, D\. Caples, A\. Draguns, N\. Ravi, P\. Putta, N\. Garg, T\. Abraham, M\. Lara, F\. Lopez, J\. Liu, A\. Gundawar, P\. Hebbar, Y\. Joo, J\. Gu, C\. London, C\. A\. S\. de Witt, and S\. Motwani \(2025\)REAL: benchmarking autonomous agents on deterministic simulations of real websites\.CoRRabs/2504\.11543\.External Links:[Link](https://doi.org/10.48550/arXiv.2504.11543),[Document](https://dx.doi.org/10.48550/ARXIV.2504.11543),2504\.11543Cited by:[§2\.1](https://arxiv.org/html/2606.06708#S2.SS1.p1.1)\.
- L\. K\. Jang, J\. Y\. Koh, D\. Fried, and R\. Salakhutdinov \(2026\)Odysseys: benchmarking web agents on realistic long horizon tasks\.External Links:2604\.24964,[Link](https://arxiv.org/abs/2604.24964)Cited by:[§2\.1](https://arxiv.org/html/2606.06708#S2.SS1.p1.1)\.
- M\. Kang, W\. Chen, D\. Han, H\. A\. Inan, L\. Wutschitz, Y\. Chen, R\. Sim, and S\. Rajmohan \(2025\)ACON: optimizing context compression for long\-horizon LLM agents\.CoRRabs/2510\.00615\.External Links:[Link](https://doi.org/10.48550/arXiv.2510.00615),[Document](https://dx.doi.org/10.48550/ARXIV.2510.00615),2510\.00615Cited by:[§2\.2](https://arxiv.org/html/2606.06708#S2.SS2.p1.1)\.
- S\. Kara, F\. E\. Faisal, and S\. Nath \(2025\)WAREX: web agent reliability evaluation on existing benchmarks\.CoRRabs/2510\.03285\.External Links:[Link](https://doi.org/10.48550/arXiv.2510.03285),[Document](https://dx.doi.org/10.48550/ARXIV.2510.03285),2510\.03285Cited by:[§2\.3](https://arxiv.org/html/2606.06708#S2.SS3.p1.1)\.
- I\. Kerboua, S\. O\. Shayegan, M\. Thakkar, X\. H\. Lù, L\. Boisvert, M\. Caccia, J\. Espinas, A\. Aussem, V\. Eglin, and A\. Lacoste \(2025a\)FocusAgent: simple yet effective ways of trimming the large context of web agents\.CoRRabs/2510\.03204\.External Links:[Link](https://doi.org/10.48550/arXiv.2510.03204),[Document](https://dx.doi.org/10.48550/ARXIV.2510.03204),2510\.03204Cited by:[§2\.2](https://arxiv.org/html/2606.06708#S2.SS2.p1.1)\.
- I\. Kerboua, S\. O\. Shayegan, M\. Thakkar, X\. H\. Lù, M\. Caccia, V\. Eglin, A\. Aussem, J\. Espinas, and A\. Lacoste \(2025b\)LineRetriever: planning\-aware observation reduction for web agents\.CoRRabs/2507\.00210\.External Links:[Link](https://doi.org/10.48550/arXiv.2507.00210),[Document](https://dx.doi.org/10.48550/ARXIV.2507.00210),2507\.00210Cited by:[§2\.2](https://arxiv.org/html/2606.06708#S2.SS2.p1.1)\.
- G\. Kim, P\. Baldi, and S\. McAleer \(2023\)Language models can solve computer tasks\.CoRRabs/2303\.17491\.External Links:[Link](https://doi.org/10.48550/arXiv.2303.17491),[Document](https://dx.doi.org/10.48550/ARXIV.2303.17491),2303\.17491Cited by:[§2\.4](https://arxiv.org/html/2606.06708#S2.SS4.p2.1)\.
- J\. Y\. Koh, R\. Lo, L\. Jang, V\. Duvvur, M\. C\. Lim, P\. Huang, G\. Neubig, S\. Zhou, R\. Salakhutdinov, and D\. Fried \(2024\)VisualWebArena: evaluating multimodal agents on realistic visual web tasks\.CoRRabs/2401\.13649\.External Links:[Link](https://doi.org/10.48550/arXiv.2401.13649),[Document](https://dx.doi.org/10.48550/ARXIV.2401.13649),2401\.13649Cited by:[§2\.1](https://arxiv.org/html/2606.06708#S2.SS1.p1.1)\.
- I\. Levy, B\. Wiesel, S\. Marreed, A\. Oved, A\. Yaeli, and S\. Shlomov \(2024\)ST\-webagentbench: A benchmark for evaluating safety and trustworthiness in web agents\.CoRRabs/2410\.06703\.External Links:[Link](https://doi.org/10.48550/arXiv.2410.06703),[Document](https://dx.doi.org/10.48550/ARXIV.2410.06703),2410\.06703Cited by:[§2\.3](https://arxiv.org/html/2606.06708#S2.SS3.p3.1)\.
- X\. Li, S\. Yu, M\. Pan, Y\. Sun, B\. Li, D\. Song, X\. Lin, and W\. Shi \(2026\)Unsafer in many turns: benchmarking and defending multi\-turn safety risks in tool\-using agents\.CoRRabs/2602\.13379\.External Links:[Link](https://doi.org/10.48550/arXiv.2602.13379),[Document](https://dx.doi.org/10.48550/ARXIV.2602.13379),2602\.13379Cited by:[§2\.3](https://arxiv.org/html/2606.06708#S2.SS3.p2.1)\.
- Z\. Lu, T\. Weng, Y\. Yang, Y\. Zhao, X\. Huang, and W\. Jiang \(2026\)BEAP\-agent: backtrackable execution and adaptive planning for GUI agents\.CoRRabs/2601\.21352\.External Links:[Link](https://doi.org/10.48550/arXiv.2601.21352),[Document](https://dx.doi.org/10.48550/ARXIV.2601.21352),2601\.21352Cited by:[§2\.4](https://arxiv.org/html/2606.06708#S2.SS4.p2.1)\.
- K\. Ma, H\. Zhang, H\. Wang, X\. Pan, and D\. Yu \(2023\)LASER: LLM agent with state\-space exploration for web navigation\.CoRRabs/2309\.08172\.External Links:[Link](https://doi.org/10.48550/arXiv.2309.08172),[Document](https://dx.doi.org/10.48550/ARXIV.2309.08172),2309\.08172Cited by:[§2\.4](https://arxiv.org/html/2606.06708#S2.SS4.p2.1)\.
- A\. Menon, M\. Saebo, T\. Crosse, S\. Gibson, E\. Jang, and D\. Cruz \(2026\)Inherited goal drift: contextual pressure can undermine agentic goals\.CoRRabs/2603\.03258\.External Links:[Link](https://doi.org/10.48550/arXiv.2603.03258),[Document](https://dx.doi.org/10.48550/ARXIV.2603.03258),2603\.03258Cited by:[§1](https://arxiv.org/html/2606.06708#S1.p4.1)\.
- S\. Schmidgall and M\. Moor \(2025\)AgentRxiv: towards collaborative autonomous research\.CoRRabs/2503\.18102\.External Links:[Link](https://doi.org/10.48550/arXiv.2503.18102),[Document](https://dx.doi.org/10.48550/ARXIV.2503.18102),2503\.18102Cited by:[§2\.3](https://arxiv.org/html/2606.06708#S2.SS3.p3.1)\.
- Y\. Tan, Z\. Gao, and X\. Wu \(2026\)Enhancing web agents with a hierarchical memory tree\.CoRRabs/2603\.07024\.External Links:[Link](https://doi.org/10.48550/arXiv.2603.07024),[Document](https://dx.doi.org/10.48550/ARXIV.2603.07024),2603\.07024Cited by:[§2\.2](https://arxiv.org/html/2606.06708#S2.SS2.p1.1)\.
- H\. Trivedi, T\. Khot, M\. Hartmann, R\. Manku, V\. Dong, E\. Li, S\. Gupta, A\. Sabharwal, and N\. Balasubramanian \(2024\)AppWorld: A controllable world of apps and people for benchmarking interactive coding agents\.CoRRabs/2407\.18901\.External Links:[Link](https://doi.org/10.48550/arXiv.2407.18901),[Document](https://dx.doi.org/10.48550/ARXIV.2407.18901),2407\.18901Cited by:[§2\.2](https://arxiv.org/html/2606.06708#S2.SS2.p1.1)\.
- T\. Wang, S\. Gooding, F\. Hartmann, O\. Riva, and E\. Grefenstette \(2026a\)A subgoal\-driven framework for improving long\-horizon LLM agents\.CoRRabs/2603\.19685\.External Links:[Link](https://doi.org/10.48550/arXiv.2603.19685),[Document](https://dx.doi.org/10.48550/ARXIV.2603.19685),2603\.19685Cited by:[§1](https://arxiv.org/html/2606.06708#S1.p1.1),[§2\.1](https://arxiv.org/html/2606.06708#S2.SS1.p2.1),[§2\.4](https://arxiv.org/html/2606.06708#S2.SS4.p2.1)\.
- X\. J\. Wang, H\. Bai, Y\. Sun, H\. Wang, S\. Zhang, W\. Hu, M\. Schroder, B\. Mutlu, D\. Song, and R\. D\. Nowak \(2026b\)The long\-horizon task mirage? diagnosing where and why agentic systems break\.CoRRabs/2604\.11978\.External Links:[Link](https://doi.org/10.48550/arXiv.2604.11978),[Document](https://dx.doi.org/10.48550/ARXIV.2604.11978),2604\.11978Cited by:[§2\.1](https://arxiv.org/html/2606.06708#S2.SS1.p2.1)\.
- Z\. Wang, K\. Wang, Q\. Wang, P\. Zhang, L\. Li, Z\. Yang, X\. Jin, K\. Yu, M\. N\. Nguyen, L\. Liu, E\. Gottlieb, Y\. Lu, K\. Cho, J\. Wu, L\. Fei\-Fei, L\. Wang, Y\. Choi, and M\. Li \(2025\)RAGEN: understanding self\-evolution in LLM agents via multi\-turn reinforcement learning\.CoRRabs/2504\.20073\.External Links:[Link](https://doi.org/10.48550/arXiv.2504.20073),[Document](https://dx.doi.org/10.48550/ARXIV.2504.20073),2504\.20073Cited by:[§2\.4](https://arxiv.org/html/2606.06708#S2.SS4.p2.1)\.
- Z\. Wang, Y\. Cui, L\. Zhong, Z\. Zhang, D\. Yin, B\. Y\. Lin, and J\. Shang \(2024\)OfficeBench: benchmarking language agents across multiple applications for office automation\.CoRRabs/2407\.19056\.External Links:[Link](https://doi.org/10.48550/arXiv.2407.19056),[Document](https://dx.doi.org/10.48550/ARXIV.2407.19056),2407\.19056Cited by:[§2\.2](https://arxiv.org/html/2606.06708#S2.SS2.p1.1)\.
- Y\. Wu, Y\. Zheng, T\. Xu, Z\. Zhang, Y\. Yu, J\. Zhu, C\. Ma, B\. Lin, B\. Dong, H\. Zhu, R\. Huang, and G\. Yu \(2026\)ContextBudget: budget\-aware context management for long\-horizon search agents\.CoRRabs/2604\.01664\.External Links:[Link](https://doi.org/10.48550/arXiv.2604.01664),[Document](https://dx.doi.org/10.48550/ARXIV.2604.01664),2604\.01664Cited by:[§2\.2](https://arxiv.org/html/2606.06708#S2.SS2.p1.1)\.
- T\. Xie, D\. Zhang, J\. Chen, X\. Li, S\. Zhao, R\. Cao, T\. J\. Hua, Z\. Cheng, D\. Shin, F\. Lei, Y\. Liu, Y\. Xu, S\. Zhou, S\. Savarese, C\. Xiong, V\. Zhong, and T\. Yu \(2024\)OSWorld: benchmarking multimodal agents for open\-ended tasks in real computer environments\.CoRRabs/2404\.07972\.External Links:[Link](https://doi.org/10.48550/arXiv.2404.07972),[Document](https://dx.doi.org/10.48550/ARXIV.2404.07972),2404\.07972Cited by:[§2\.1](https://arxiv.org/html/2606.06708#S2.SS1.p1.1)\.
- T\. Xue, W\. Qi, T\. Shi, C\. H\. Song, B\. Gou, D\. Song, H\. Sun, and Y\. Su \(2025\)An illusion of progress? assessing the current state of web agents\.CoRRabs/2504\.01382\.External Links:[Link](https://doi.org/10.48550/arXiv.2504.01382),[Document](https://dx.doi.org/10.48550/ARXIV.2504.01382),2504\.01382Cited by:[§2\.1](https://arxiv.org/html/2606.06708#S2.SS1.p1.1)\.
- D\. Yan, H\. Zhang, G\. Huzhang, Y\. Li, Y\. Wang, Q\. Chen, Z\. Xu, W\. Luo, Y\. Li, W\. Dong, and C\. Shen \(2026\)M2\{\}^\{\\mbox\{2\}\}: dual\-memory augmentation for long\-horizon web agents via trajectory summarization and insight retrieval\.CoRRabs/2603\.00503\.External Links:[Link](https://doi.org/10.48550/arXiv.2603.00503),[Document](https://dx.doi.org/10.48550/ARXIV.2603.00503),2603\.00503Cited by:[§2\.2](https://arxiv.org/html/2606.06708#S2.SS2.p1.1)\.
- R\. Ye, Z\. Zhang, K\. Li, H\. Yin, Z\. Tao, Y\. Zhao, L\. Su, L\. Zhang, Z\. Qiao, X\. Wang, P\. Xie, F\. Huang, S\. Chen, J\. Zhou, and Y\. Jiang \(2025\)AgentFold: long\-horizon web agents with proactive context management\.CoRRabs/2510\.24699\.External Links:[Link](https://doi.org/10.48550/arXiv.2510.24699),[Document](https://dx.doi.org/10.48550/ARXIV.2510.24699),2510\.24699Cited by:[§1](https://arxiv.org/html/2606.06708#S1.p5.1),[§2\.2](https://arxiv.org/html/2606.06708#S2.SS2.p1.1)\.
- H\. Yen, A\. Paranjape, M\. Xia, T\. Venkatesh, J\. Hessel, D\. Chen, and Y\. Zhang \(2025\)Lost in the maze: overcoming context limitations in long\-horizon agentic search\.External Links:2510\.18939,[Link](https://arxiv.org/abs/2510.18939)Cited by:[§2\.2](https://arxiv.org/html/2606.06708#S2.SS2.p1.1)\.
- A\. L\. Zhang, T\. Kraska, and O\. Khattab \(2025a\)Recursive language models\.CoRRabs/2512\.24601\.External Links:[Link](https://doi.org/10.48550/arXiv.2512.24601),[Document](https://dx.doi.org/10.48550/ARXIV.2512.24601),2512\.24601Cited by:[§1](https://arxiv.org/html/2606.06708#S1.p6.1),[§2\.4](https://arxiv.org/html/2606.06708#S2.SS4.p1.1)\.
- G\. Zhang, J\. Wang, J\. Chen, W\. Zhou, K\. Wang, and S\. Yan \(2025b\)AgenTracer: who is inducing failure in the LLM agentic systems?\.CoRRabs/2509\.03312\.External Links:[Link](https://doi.org/10.48550/arXiv.2509.03312),[Document](https://dx.doi.org/10.48550/ARXIV.2509.03312),2509\.03312Cited by:[§2\.3](https://arxiv.org/html/2606.06708#S2.SS3.p3.1)\.
- Y\. Zhang, T\. Yu, and D\. Yang \(2024\)Attacking vision\-language computer agents via pop\-ups\.CoRRabs/2411\.02391\.External Links:[Link](https://doi.org/10.48550/arXiv.2411.02391),[Document](https://dx.doi.org/10.48550/ARXIV.2411.02391),2411\.02391Cited by:[§2\.3](https://arxiv.org/html/2606.06708#S2.SS3.p1.1)\.
- S\. Zhou, F\. F\. Xu, H\. Zhu, X\. Zhou, R\. Lo, A\. Sridhar, X\. Cheng, Y\. Bisk, D\. Fried, U\. Alon, and G\. Neubig \(2023\)WebArena: A realistic web environment for building autonomous agents\.CoRRabs/2307\.13854\.External Links:[Link](https://doi.org/10.48550/arXiv.2307.13854),[Document](https://dx.doi.org/10.48550/ARXIV.2307.13854),2307\.13854Cited by:[§1](https://arxiv.org/html/2606.06708#S1.p1.1),[§2\.1](https://arxiv.org/html/2606.06708#S2.SS1.p1.1)\.