HarnessMM paper cuts agent token use 87.9% on DeepSeek harness by reusing decisions
blaizedsouza · x · 2026-10-26
An arXiv paper introduces HarnessMM, which caches a coding agent's resolved decisions (exact actions, literal arguments, guards) in Executable Decision Graphs and replays them at the model-tool boundary instead of re-invoking reasoning after every tool result. With recursive self-improvement refining harness configs while actor weights stay frozen, a DeepSeek harness saw tokens/task fall from 96.28K to 11.62K (-87.93%) while success rose from 71.43% to 85.71%; Codex on HumanEval kept 90.82% success with 29.54% token savings. Key insight: much of what we pay for is the model re-deciding what it already decided.
Related event: HarnessMM Cuts Agent Harness Token Use by 88%(2 posts)→
More from coding & agent
- Dev mixes Claude specs with abliterated GLM to build SS7 threat detector — l4rz · 2026-10-27
- Cloudflare Durable Objects validated as top architecture for AI agent WebSocket servers — tristanbob · 2026-10-27
- GitHub Next's Chopin prototype explores real-time multiplayer agentic planning — max__drake · 2026-10-27
- TMLR paper: enforced thinking doesn't reliably make LLM agents disclose key info to users — SharonYixuanLi · 2026-10-27
- LLM-written Rust library bbtree verified: speed claims hold, memory savings reverse — lmoroney · 2026-10-27
- Resin lets an agent build standalone LLM inference engines; vLLM and SGLang optimize DeepSeek kernels differently — AccBalanced · 2026-10-27