HarnessMM 通过可执行决策图缓存智能体框架已确定的决策，在 DeepSeek 框架上将每任务 token 从 96.28k 降至 11.62k，成功率从 71.43% 升至 85.71%。该框架用递归自我改进（RSI）循环根据训练反馈调优框架配置，模型权重不变。在 Codex 上跑 HumanEval，成功率保持 90.82%，每任务 token 下降 29.54%。
Learn to build a custom harness, folks!
There is so much alpha in understanding how to optimize the harness.
This work shows the benefits of a smart harness configuration and caching the decisions your agent harness has already made.
(bookmark it)
They propose HarnessMM, which cuts tokens per task on a DeepSeek harness from 96.28k to 11.62k while success rises from 71.43% to 85.71%.
Harnesses often call the model again after tool feedback even when an earlier call already settled the next action. HarnessMM stores those settled decisions in an Executable Decision Graph with exact actions, arguments, dependencies, and guards. When the guard holds, the controller reuses the decision. Anything new goes back to the model.
A recursive self-improvement (RSI) loop tunes the harness configuration from training feedback, and the model weights never change.
With Codex on HumanEval, success stays at 90.82%, and tokens per task drop 29.54%.
Before you pay for a bigger model, check how many of your harness calls repeat a decision the agent already made.
Paper: https://arxiv.org/abs/2610.32813
Chat with Paper: https://academy.dair.ai/papers/harnessmm-recursive-self-improvement-for-harness-token-efficiency-2610.32813
来源：elvis · x.com