Brief, mostly signal

Google Research releases Harness2 for recursive agent harnessing

The framework refines prompts, tools and memory with model weights fixed. It reports gains up to 11.5 points on WorkBuddy-Bench and needs Linux and Python 3.12.

Read the source at github.com Discussion

What the research found

Google Research's Harness2 repo describes recursive harness optimization with weights fixed, reporting gains up to 11.5 points on WorkBuddy-Bench for GLM-5 and Gemini 3.5 Flash; needs Linux, Python 3.12 and uv.

Claims and how well the evidence supports them

Harness2 keeps model weights fixed while refining harness components 98%
Gains of up to 11.5 points on WorkBuddy-Bench for GLM-5 and Gemini 3.5 Flash 97%
Setup requires Linux, Python 3.12 and uv 93%
Model-backed benchmarks need Google Cloud CLI, Vertex AI config and Docker for WorkBuddy-Bench 86%

Evidence

github.com https://github.com/google-research/harness2
news.ycombinator.com https://news.ycombinator.com/item?id=50151059

Actionability rated medium.