HarnessMM cuts DeepSeek Harness tokens by 87.93% while raising success from 71.43% to 85.71%
rohanpaul_ai · x · 2026-11-02
A Tsinghua and Microsoft paper introduces HarnessMM, boosting agent harness token efficiency via guarded decision reuse and recursive self-improvement (RSI). Key insight: harnesses re-invoke the LLM after every tool result even when the next step is already determined (e.g., run tests once code compiles).
Method
- Executable Decision Graphs store exact actions, literal arguments, causal dependencies, and public guards
- A controller at the model-tool boundary reuses resolved decisions to replace later reasoning callbacks; actor weights stay frozen
- Certified Callback Substitution executes decisions through the native interface; new semantic decisions, weak evidence, or selector abstention fall back to the original actor
- RSI iteratively refines harness configurations from training feedback, with validation picking the deployment config
Results
- Codex: success preserved at 90.82%, tokens/task drop from 66.72k to 47.02k (29.54% saving)
- DeepSeek Harness: success rises 71.43% → 85.71%, tokens/task 96.28k → 11.62k (87.93% saving)
- MBPP ablation: 26.91% cheaper than single-use graphs
More from coding & agent
- InfoQ: A Systemic Approach to Memory, Knowledge, and Context in Agentic AI Architectures — AdiPolak · 2026-11-02
- Autonomous's Harness Pro uses an ESP32-P4, not a mini PC, to monitor AI coding agents — dee_hw · 2026-11-02
- RL-trained LFM2.5-24B adapter matches accuracy with 20% fewer tool calls — yb2698 · 2026-11-02
- Debugging production AI agents: error surfaces at step 5, root cause at step 3 — Ok-Machine7983 · 2026-11-02
- RAGFlow 1.0 preview: from RAG engine to agent-native database — newpeak · 2026-11-02
- Kitsune: open-source MCP server that shares your Firefox with a coding agent — Sad-Interview9085 · 2026-11-02