Google open-sources Harness² for self-improving AI agents

Google Research has released Harness² under Apache 2.0, a framework that lets agents rewrite their prompts, tools, memory and scripts through two recursive loops without retraining the model. A single step (k=1) improved GLM-5 and Gemini 3.5 Flash on three benchmarks, including the 1,143-task LAB.

• Harness² code, paper and demo are released under Apache 2.0
• Agents edit eight harness components while model weights stay fixed
• Improvements come from contrasting runs, with no ground-truth answers
• At k=1, GLM-5 and Gemini 3.5 Flash gained on three benchmarks