Same Problem, 8x Lower Cost! We Built the Most Cost-Efficient Harness, Maka

Related posts: https://x.com/AstroHanRay/status/2083918530605707459?s=20

https://x.com/jakevin7/status/2078726365894021482?s=20

https://x.com/jakevin7/status/2072923081463763342?s=20

On the same problem, with the same DeepSeek, a conventional Harness costs $100—we get it done for $12.50, and our pass rate is more than ten percentage points higher. Let’s break down the results: how the open-source Agent Harness “Maka” uses an event-log architecture to achieve three things at once—the lowest cost, the fewest tokens, and the best fit for Chinese models. Every one of them has a code-level mechanism, and every one of them is backed by public benchmark scores.

What We Did

• Open-sourced the Agent Harness “Maka”: local-first, Apache-2.0, TypeScript.

• One Runtime covers four endpoints—Desktop / TUI / CLI / Headless—with the same model connections, permissions, and sessions shared across endpoints.

• Same DeepSeek V4 Flash, one problem: cost is 1/8 of OpenCode’s.

• Terminal-Bench 2.1: $0.03215 per problem passed, the lowest among the four arms (Codex / CC / Reasonix).

• Same K3 model: Maka 69.7% vs official KimiCode 59.6%.

Project: https://github.com/maka-agent/maka-agent

2.5 months, 1,240+ stars, 49 contributors, 83 PRs merged within 23 hours of the v0.1.4 release. The 200-person community group is full; group 2 is being opened.

1. Where Agent Costs Burn: Four Bottlenecks, Each a Harness Optimization Opportunity

Model API pricing keeps dropping year over year, but the bill for running Agents hasn’t followed. The money burns in four places, and every one of them falls under the Harness’s control:

• Repeated re-derivation of thinking. Reasoning the model already did in the previous step gets re-done from scratch in the next step.

• Repeated reads of tool results. The same code, the same output—the model reads them over and over.

• Redundant prompt and tool surface. A 20KB product prompt plus a full tool surface means paying for irrelevant content on every turn.

• Low cache hit rate. Full history gets re-transmitted every time, so the provider’s cache discounts are never captured.

Maka’s cost savings come from a design that solves all four bottlenecks at once. Design first, then mechanisms.

2. Technical Foundation: Four Design Tenets

All of Maka’s engineering decisions converge into four tenets, written in the architecture docs. Understand them, and you’ll understand every number that follows.

1. Log is the Runtime

Model messages, tool calls, tool results, and termination reasons all land in the Runtime Event Log. Sessions, UI, model context, and crash recovery are all projections of this log.

Desktop / TUI / Headless ↓ SessionManager → AgentRun → Model + Tool Runtime ↓ Runtime Event Log → Context / Session / UI projections ↓ Task Event Log → TaskRun → Self-check / AHE

The benefit lands directly in engineering: crash recovery rebuilds execution state from the log; benchmarks, failure postmortems, and token usage can all be traced back from the log. One log drives all four entry points.

2. Context is not history

What the model sees next is decided by the Context layer; recorded execution facts stay in the event log. Save context when it should be saved, and read it back when lookback is needed.

This tenet directly gave rise to the Context-budget pruning below.

3. A task may outlive a Turn

Headless uses TaskRun + Task Event Log + budget + continuation to support interruptible, inspectable, resumable persistent tasks. If power or network drops mid-run, it picks back up.

4. Feedback is not fact authority

Self-check produces evidence and grants one bounded repair opportunity. “I checked it” does not automatically upgrade to system fact. This keeps the agent’s self-confirmation bias from polluting the execution record.

3. Most Cost-Efficient: Three Mechanisms, Down to the Implementation Layer

Mechanism 1: Thinking Replay Protocol—Eliminating Duplicate Reasoning

While running benchmarks, we dug a protocol-level problem out of the telemetry. It directly explained why early Maka burned more tokens than Codex.

Maka’s model adapter layer has three switches in runtimeEventReplaySupport() at packages/runtime/src/model-adapter.ts:141:

• signedThinking: Anthropic only

• unsignedThinking: Kimi only

• openAiResponsesThinking: Responses API only

DeepSeek takes the openai-chat path, where all three switches are false—each step strips the model’s past thinking from the prompt. The consequence shows up monotonically in telemetry:

At every step, the model re-derives “what it already thought in the previous step”; the longer the history, the more expensive the re-derivation.

The fix: a unified ReasoningReplayContract, with DeepSeek moved to the Responses API to retain thinking. After the fix, the team’s own words: “way faster.”

Kimi’s official K3 blog said the same thing: if the harness doesn’t pass thinking back completely, generation quality becomes severely unstable. A one-line switch at the protocol layer directly decides both the result and the bill.

Mechanism 2: Context-budget Pruning—Eliminating Repeated Reads

Tool output is often longer than model input. Maka’s Context layer prunes: tool results exceeding 2048 tokens are replaced with a placeholder and archived; the model reads them back only when needed.

A/B test across 121 Terminal-Bench tasks:

41.7% token savings, and performance went up rather than down.

Mechanism 3: Slimmed System Prompt + Slimmed Tool Surface

In benchmarks, Maka’s system prompt is only 253 bytes, 4 lines; KimiCode uses an ~20KB product prompt plus a full tool surface. The foreground only exposes six tools: Read / Write / Edit / Bash / Glob / Grep.

Thousands of tokens of prompt overhead are saved each turn, and the model’s attention isn’t diluted by irrelevant tools.

Data Recap

• Same DeepSeek, one problem: cost is roughly 1/8 of OpenCode’s.

• In the four-arm benchmark, Maka uses 61% of Codex’s input tokens to get 93% of its score.

• Input cache hit rate: 97.93% (DeepSeek’s cache-hit input price is 2% of a cache miss).

4. Best Fit for Chinese Models: Five Technical Evidence Points

Evidence 1: Same K3, 10.1 Points Ahead of the Official Harness

Same K3 model, max thinking setting, thinking preserved at both ends. The gap comes from Maka’s full-replay