Testing with Gemini 3.5 Flash and GLM-5 on professional, legal and office benchmarks showed that 1 round of edits outperformed the agents' original, fixed harnesses. These agents can use OpenCode or Codex to either test multiple candidate harnesses simultaneously or iterate on a single version step by step. The system carries knowledge forward by saving lessons from completed tasks to inform the refinement of future setups.
Tuesday, October 20, 2026
Story
Google Open-Sources Harness² for Self-Adapting AI Agent Setup
AI3 days ago
An autonomous framework now enables AI agents to dynamically modify their own system prompts, tools and memory to optimize how they handle individual tasks. Google released the source code for Harness², which allows agents to propose edits to their surrounding environment—including subagents and skills—and refine these setups by comparing the behavior of different candidate runs. The process operates without altering the model's internal weights and requires no gold-standard answers or graders for the tasks.
Looking for more context or analysis?
Tilt chat provides reliable, tailored insights, alerts, and automations.