Google open-sources Harness², letting agents self-improve prompts, tools and memory with frozen weights
CurieuxExplorer · x · 2026-10-19
Google has open-sourced Harness², a framework that turns any agent harness into a self-improving system.
- The harness is everything wrapped around the model: system prompt, tools, memory, skills, subagents. Traditionally a human designs it once and it stays fixed.
- Harness² lets the agent refine that setup per task while model weights stay frozen: it proposes harness edits, runs candidate versions, and composes a better harness by comparing how those runs behaved.
- Because it learns from run-to-run differences, it needs no ground truth or grader — each task leaves evidence that makes the next refinement smarter.
The takeaway: agentic systems can recursively improve their own harness, so the model is no longer the bottleneck.
More from coding & agent
- 'Vibe Coding' Has Lost Its Meaning: From Midnight Hacks to 500-Hour Agent Projects — chrisalbon · 2026-10-19
- Next Token Podcast #006: Grok Bot & Muse Hands-On, AI Reverse Engineering, Agent Pitfalls — vista8 · 2026-10-19
- Image Lab 2 rebuilds AI image workflow: AI acts as creative director before generating — woulatte · 2026-10-19
- Biscotti: a self-hosted Instinct-like iMessage assistant in ~850 LOC — lucasgelfond · 2026-10-19
- CS Freshman Money Guide: Ship a Real Tool First, Upgrade Gear Later — sven_ai · 2026-10-19
- Green Room: a chat room built for AI agents to talk to each other — Special_Reindeer7685 · 2026-10-19