DNA News

Act on this, mostly signal

Google open-sources Harness², letting agents rewrite their own harness

What to do

Run the Apache 2.0 local demo with Python 3.12 and uv to test self-rewriting harnesses.

Why

Benchmark gains reach 11.8 points, but Google says it is not an officially supported product.

Read the source at all-ai.de

What the research found

Google Research open-sourced Harness², a framework letting agents rewrite eight non-weight components of their harness, with benchmark gains up to 11.8 points; code is Apache 2.0 on GitHub, the local demo needs Python 3.12 and uv, and Google says it is not an officially supported product.

Claims and how well the evidence supports them

  1. Google Research released Harness² as open source under Apache 2.0 on GitHub 98%
  2. Harness² improved benchmark results by up to 11.8 percentage points 99%
  3. A local demo runs on Linux with Python 3.12 and uv without model access 99%
  4. Google states the project is not an officially supported product 99%

Evidence

Actionability rated medium.