HarnessMM cuts agent tokens with cached decisions

// 1h agoRESEARCH PAPER

HarnessMM cuts agent tokens with cached decisions

HarnessMM is a research method that reuses validated tool decisions through guarded decision graphs, reducing repeated reasoning while keeping the underlying model frozen. The paper reports 29.54% lower tokens per task with Codex and 87.93% lower tokens with DeepSeek Harness.

// ANALYSIS

HarnessMM makes a compelling case that agent efficiency is increasingly a harness problem, not just a model problem.

–Reuses exact actions only when observed tool receipts satisfy predefined guards
–Falls back to the original model for new or ambiguous decisions
–Preserves Codex success on HumanEval while lowering token usage
–Reports substantial gains on DeepSeek Harness, though broader validation is still needed
–Recursive harness optimization could become a practical alternative to retraining for improving agent cost and reliability

// TAGS

harnessmmagenttool-useinferencecontext-engineeringresearch

DISCOVERED

1h ago

2026-10-26

PUBLISHED

1h ago

2026-10-26

RELEVANCE

9/ 10

AUTHOR

omarsar0

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

VIDEO30m ago

Claude Opus 5.5 Creates Unreal Cities

A demo from Shikhar shows Claude Opus 5.5 generating an entire city in Unreal Engine, with Blender used to create all assets. It is a striking proof of accelerated 3D prototyping, not evidence that one prompt can ship a million-dollar game.

INFRA38m ago

mlx-audio Brings Fast Local Audio Inference to Macs

mlx-audio is an open-source Apple MLX library for running TTS, STT, speech-to-speech, and music-generation models locally on Apple Silicon. Its Python, CLI, and OpenAI-compatible API make local audio inference easier to integrate into developer workflows.

OPEN SOURCE48m ago

dithercn Launches Agent-Built Dithered UI

dithercn gives coding agents a WebGL dither engine, agent skill, and 52 React/shadcn examples for building dithered backgrounds, charts, interface elements, text, and images. Components install as editable source through the shadcn CLI.