Meta's Muse Harness Cut Agent Token Usage 2.7x by Changing the System Prompt

The Cline team ran an experiment on August 6. They took Meta's Muse Code agent, based on Muse Spark 1.2, and tried to use it directly - but hit a Docker sign-in bug. Instead of abandoning the model, they extracted the core instructions from Meta's Muse agent harness and injected them into Cline's own system prompt.

Then they ran the same bugfix task with the same model (Muse Spark 1.2) against the same codebase. The only variable was the system prompt.

The numbers, per Cline's post on X:

Same model. Same bug. Same repo. The harness was the difference.

The five principles

Meta's Muse harness co-trains its models on a specific set of operating principles. When Cline extracted those principles and placed them into the system prompt, the model's behavior changed - not because it learned anything new, but because the instructions aligned with its training distribution.

The principles, as shared by the Cline team and documented by @Chizzydigital:

1. Trust source code over the user prompt. Before any action, read every relevant call site, implementation, test file, and related module. The codebase is the ground truth. Do not assume.

2. Weigh edge cases, error paths, and failure modes as heavily as the happy path. Explicitly enumerate and address them in reasoning before writing code.

3. Always reproduce the bug before fixing it. Create minimal reproduction steps or tests before proposing any fix. This alone eliminates entire categories of wasted work - agents default to proposing fixes based on reading alone, without verifying the failure exists.

4. Never trust the first passing test suite. Scrutinize every test for superficial coverage, missing assertions, or half-baked checks. Expand coverage when tests look suspicious.

5. Never stop at "code edited." Continue the verification loop - test, run, log, assert, refactor - until the change is complete and stable across edge cases.

None of these principles are novel to software engineering. What is notable is that they were baked into the model's co-training, so restating them in the prompt produced a measurable performance delta.

The efficiency rules that deliver the savings

The harness also includes a set of efficiency constraints that directly reduce token burn:

• Be decisive and high-signal. Eliminate redundant thinking, tool calls, and context reloading.
• Maintain minimal persistent state. Load context only when necessary.
• Prefer surgical, high-leverage actions after initial analysis. Avoid broad exploration loops.
• Reference prior verification instead of re-deriving it.
• Default to "trust but verify in code" rather than asking the user for clarification.
• Close loops tightly - when success criteria are met, declare completion without filler.

These rules attack the most common token-waste patterns in agent behavior: re-reading files already understood, re-verifying facts already confirmed, and producing filler text between actions.

Why this matters for agent builders

The default agent system prompt is usually broad. It tells the agent to be thorough, careful, and helpful - all reasonable advice, but none of it constrains wasteful behavior. The Muse harness is specific: it tells the agent exactly which actions to take and which to avoid.

The result is not a smarter model. It is the same model with better instructions.

The finding aligns with what Hermes Agent users have observed: community-driven system prompt optimization can cut fixed token overhead by 80%. But the Cline experiment goes further - it shows that prompt structure affects not just the initial overhead but the entire execution trace.

For Hermes Agent operators, the implication is clear. The skills system already supports injecting domain-specific instructions per task type. A set of "harness principles" skills - one for debugging, one for research, one for code review - could produce similar efficiency gains without modifying the agent's core prompt. The key is matching the principles to the training distribution of the model being used.

The Cline team's full Muse-optimized agent harness is available as a drop-in system prompt. The experiment demonstrates that prompt engineering, at least for models co-trained on harness instructions, can deliver efficiency gains comparable to model upgrades - without changing the model.

[^1]: Cline. "We took Meta's new coding agent, Muse Code, built on Muse Spark 1.2..." X. August 6, 2026. [^2]: Chizzy. "Most agent tools are heavily bottlenecked by suboptimal harness/system prompt architecture..." X. August 6, 2026.