Google open sources EnvHarness for self-evolving AI agents

Google has open sourced EnvHarness, a plug-in layer that dynamically adapts an agent's training environment to its weaknesses without modifying the agent, environment, or evaluation system. On WebArena, SWE-bench Verified, and ALFWorld it delivered gains of 2.9–3.7 percentage points across Gemini 3.1 Flash-Lite, Qwen3.6 27B, Gemini 3.5 Flash, and Claude Sonnet 4.6.

• EnvHarness wraps an environment via its standard reset() and step() interface
• setup, rule, and link components customize state, transitions, and observations
• EnvRigger analyzes rollouts to find weaknesses and generate targeted mutations
• Gains of 2.9–3.7 points on WebArena, SWE-bench Verified, and ALFWorld