Google releases Harness² framework for AI prompt and tool reorganization
Google Cloud AI Research and Northwestern University posted the Harness² paper on arXiv on October 9, 2026 and open‑sourced the code on GitHub (≈16,700 stars). The framework lets an AI model keep its pre‑trained weights while a recursive system rewrites eight components – system prompts, guardrails, memory, tool descriptions, sub‑agents, skills, scripts and plugins – to better suit each task. It…
Key points
- Harness² rewrites prompts, tools and other components without retraining, yielding up to 11.5‑point gains per review.
- Reported benchmark lifts include 43.1→77.8 on verification and GLM‑5 64.2→75.7, surpassing Gemini 3.5 Flash.
- Open‑source repository (≈16,700 stars) needs Linux, Python 3.12, uv; not an officially supported Google product.
Benchmarks show up to an 11.5‑point lift per review. In a verification case the score rose from 43.1 to 77.8. On the WorkBuddy‑Bench, GLM‑5 with OpenCode improved from 64.2 to 75.7, overtaking Gemini 3.5 Flash (72.4). With one review Gemini 3.5 Flash and OpenCode moved to 77.5 and reached 80.8 when five candidates were tried in parallel. The authors note that removing the notebook drops performance, confirming the learning effect. Installation requires Linux, Python 3.12, and the uv tool; a demo runs without an API key but the README warns about unrestricted network access, disabled sandboxing for Codex, and API cost accumulation. The code is labeled research‑only and not an officially supported Google product.
Model page: Jev →
The story so far
4 episodes →- Google releases Harness² framework for AI prompt and tool reorganizationthis story
Google's "Harness2" has arrived. A 5-minute breakdown of how AI improves its tools by reorganizing the kitchen without changing the chef's skills
note.com · 21 October 2026
Loading the full article…
This text was published by note.com and written by DenDen. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Research
All →- Study finds retrieval, not reading, limits RAG in long-context QA · 1 src
- Do We Need the Largest LLMs for Every Clinical Decision? · 1 src
- Researchers achieve 77.95% F1 on ambivalence detection using audio‑text cross‑attention · 1 src
- WaveCommitments: Private and Verifiable Speech-Agent Marketplaces · 1 src
- OpenAI brings back Aleksander Madry to focus on AI safety · 2 src
Comments
via GitHub Discussions