Harness²: Google's recursive agent-harnessing framework
Google Cloud AI Research open-sourced Harness² (v0.1.0), a framework that recursively refines the 8 files around a frozen model (prompts, guardrails, memory, tools, subagents, skills, scripts, plugins) without verifiers or retraining.
Why it matters
Improves agent performance on unseen tasks by tuning scaffolding instead of weights — a single contrastive step raised GLM-5 on WorkBuddy-Bench 64.2 to 75.7; works with OpenCode and Codex executors.
What you could build with it
A small dev-tools startup could wrap Harness²'s Propose–Probe–Compose loop around an existing coding agent so its prompts and guardrails auto-tune from production failures, selling fewer broken agent runs instead of bigger models.
Does it hold up?
Too early to judge — code dropped days ago with vendor-reported benchmark gains; independent reproductions are not yet available.
Built with Harness²
- Why Google's New Harness² Release Changes How I Think About AI AgentsMedium · Hands-on writeup walking through the October 2026 paper and repo, arguing agent gains now come from scaffolding, not scale.
- Harness²: Recursive Agent Harnessing for an Open Worldarxiviq · Paper summary with benchmark numbers (WorkBuddy-Bench 64.2 to 75.7 for GLM-5, JobBench 31.5 to 39.0) and caveats about test-time token costs.
- Google Open-Sources Harness2, a Recursive Agent Harnessing Frameworkmarktechpost · Technical deep-dive with full benchmark tables (GLM-5 +11.5 pts on WorkBuddy-Bench), install/run instructions, and a feature comparison vs Meta-Harness, ReasoningBank, and Harness Scaling.
- Google's Harness²: The Open-Source Framework That Helps AI Agents Improve Without Retraining Their Modelsmedium · Hands-on walkthrough showing how to clone the repo, set up the uv/Python 3.12 environment, and run the credential-free local demo that exercises both recursion modes.
- Google Research just open-sourced a recursive Agent Harness — community discussion (30 comments)linkedin · Practitioner post (Shubham Saboo) explaining the Propose-Probe-Compose loop to a builder audience, with a 30-comment discussion thread of reactions and questions from AI engineers.
Learn more
First spotted on github: source.
More AI agents
Agent ReachOpen-source CLI and SKILL.md kit that gives AI agents read/search access to 15+ internet platforms (Twitter, Reddit…agents · JEV 0.8AI Job Search — the job hunt that runs on your machineOpen-source Claude Code framework that turns job hunting into an agentic pipeline: scrape portals, score fit, tailor…agents · JEV 0.79PentAGI — fully autonomous AI agents for penetration testingPentAGI is a self-hosted, open-source system where a team of AI agents plans and executes authorized penetration tests…agents · JEV 0.77PPT Master: AI skill that generates native, editable PowerPoint decks from any documentOpen-source AI agent workflow that turns documents or topics into real, native PowerPoint decks: native shapes…agents · JEV 0.76Docker Agent: AI agent builder and runtime by Docker EngineeringA Docker CLI plugin for building, running and sharing AI agents with declarative YAML config, multi-agent…agents · JEV 0.75K-Dense-AI/scientific-agent-skills: turn any AI agent into an AI scientistK-Dense AI's open collection of 177 validated, lazily-loaded skills that turn any agent (Cursor, Claude Code, Codex…agents · JEV 0.74
Get the week's best AI launches, plus 3 ideas worth building
One email every Saturday. Ranked by traction, not hype. Free.