HarnessMM paper reuses settled agent decisions to cut tokens per task

An arXiv paper reports that reusing an agent harness's already-settled decisions cut tokens per task by 29.5% to 87.9% on HumanEval.

A paper posted to arXiv on October 23, HarnessMM: Recursive Self-Improvement for Harness Token Efficiency, targets a common agent-loop cost: harnesses call the model again after tool feedback even when an earlier call already fixed the next action. If you pay per token for agent loops, it is a concrete idea to check against your own logs, but the evidence so far is one paper.

According to the abstract, HarnessMM stores resolved decisions in Executable Decision Graphs. These hold exact actions, literal arguments, causal dependencies and guards. A controller at the model-tool boundary reuses those decisions instead of calling the model again. New decisions, unsupported evidence or selector abstention send the step back to the original model. A recursive self-improvement loop tunes the harness configuration from training feedback, and model weights stay fixed.

The reported results are all on coding benchmarks. On HumanEval, Codex success holds at 90.82% while tokens per task fall from 66.72k to 47.02k, a 29.54% saving. A DeepSeek harness goes from 71.43% to 85.71% success and from 96.28k to 11.62k tokens per task. A separate MBPP ablation reports 26.91% lower cost than single-use graphs. @omarsar0 highlighted the work with the advice to check how many harness calls repeat a decision before paying for a bigger model.

Several points are still unclear: whether code is available, whether it plugs into existing coding agents, and whether the baselines had prompt caching on. Prompt caching discounts repeated input prefixes, while HarnessMM aims to skip reasoning calls entirely, so the savings may or may not stack. For a baseline today, Claude Code's /usage reports prompt cache statistics, including the share of input tokens served from cache.

Sources