Google's Harness² recursively self-improves agent harnesses, up to 11.8% gain without weight updates
burny_tech · x · 2026-10-21
Google's Harness² (Recursive Agent Harnessing for an Open World) tackles two pain points: harness optimization struggles to adapt to new tasks without ground-truth feedback, and improvements made for one task rarely transfer to others.
It introduces two levels of recursive self-improvement:
- For the current task, it proposes, tests, and combines changes to the agent's prompts, tools, and workflows
- It stores lessons from those experiments to improve future harness modifications
Across GLM-5 and Gemini 3.5 Flash, it improves performance by up to 11.5% on WorkBuddy-Bench, 4.5% on JobBench, and 11.8% on Legal Agent Bench — without updating model weights.
Related event: Google open-sources Harness² for recursive agent self-improvement(4 posts)→
More from coding & agent
- Building WhatsApp agents for clients: who handles the human handoff? — flavia-f · 2026-10-21
- Too many MCP tools: internal data says 77% of agent tokens are wasted — Jazzlike-Form9669 · 2026-10-21
- Agents and software companies can coexist — the bar is trust, not code quality — andreisavu · 2026-10-21
- Buying software means paying for a team that decides how it improves — can agents replace it? — andreisavu · 2026-10-21
- Does opencode phone home even with local models? Users flag data concerns — Unusual_Principle536 · 2026-10-21
- QAMCP: open-source CLI to test MCP servers and spot hidden prompt injections — abhiramreddykalluru · 2026-10-21