Google Open Sources Harness2 for Self-Improving AI Agents

·2026.10.19 07:33

Key point

Tested on Gemini 3.5 Flash and GLM-5, the system improves performance on legal and office benchmarks without changing model weights.

Google released Harness2 as an open-source system that allows AI agents to refine their own harnesses (system prompts, tools, memory, skills) for each task.
The agent proposes changes, runs candidate versions, and selects the best performer by comparing outcomes, requiring no correct answers or external graders.
Model weights remain unchanged; the system saves improved harnesses and lessons after each task to inform future edits.
Testing with Gemini 3.5 Flash and GLM-5 on legal, office, and professional benchmarks showed that even a single round of edits outperformed the original fixed harnesses, with further rounds increasing scores.

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.