Back to AI Research

AI Research

HarnessMM reuses guarded decisions to reduce repeated agent reasoning

Key Takeaways

  • Fixed actor weights and explicit fallback rules let a harness replace callbacks whose next action is already resolved.
  • An agent may call its reasoning model again after tool feedback even when an earlier step already determined the next action.
  • The [HarnessMM paper](https://arxiv.org/abs/2610.32813) studies how to remove that repeated work while preserving an explicit boundary for decisions that still need the actor.
  • ## Store the action and the conditions that justify it
  • Executable Decision Graphs retain exact actions, literal arguments, causal dependencies and public guards.

An agent may call its reasoning model again after tool feedback even when an earlier step already determined the next action. The HarnessMM paper studies how to remove that repeated work while preserving an explicit boundary for decisions that still need the actor.

Store the action and the conditions that justify it

Executable Decision Graphs retain exact actions, literal arguments, causal dependencies and public guards. A controller at the model-tool boundary activates resolved decisions using observed receipts. Certified Callback Substitution executes the action through the native interface and records removal of the registered slow callback.
New semantic decisions, unsupported evidence and selector abstention return control to the original actor. The method's claim is therefore about guarded reuse of resolved decisions, rather than substituting a cached response for any superficially similar situation.

Improve the harness while holding actor weights fixed

Recursive self-improvement refines companion harness configurations using training feedback, with validation selecting the deployment configuration. Actor weights remain fixed while the harness changes. The representation also restricts what feedback may alter.
On HumanEval, the paper reports Codex success of 90.82% while tokens per task fall from 66.72 thousand to 47.02 thousand, a 29.54% saving. For the evaluated DeepSeek Harness, success rises from 71.43% to 85.71% while tokens per task fall from 96.28 thousand to 11.62 thousand.

Judge the complete task, including fallbacks

These results belong to the stated harnesses and benchmarks. They do not establish the same savings for arbitrary tools, workflows or models, and token reductions are not automatically an equal reduction in end-to-end latency.
A useful reproduction should inspect whether a reused action still satisfies its guards after the actual tool result. Measure task success and cost together, including cases where the selector abstains and the actor resumes. The study's practical question is which reasoning calls can be removed because their decision is already resolved, without concealing a genuinely new decision behind the optimization.

Comments