Google open-sources Harness², reporting gains of up to 11.5 points
Google Cloud AI Research has released Harness², an open-source framework that changes an AI agent’s surrounding setup—such as prompts, tools and memory—while leaving the underlying model weights fixed. In tests across three professional benchmarks, the researchers reported gains of up to 11.5 points on WorkBuddy-Bench, 4.5 points on JobBench and 11.8 percentage points on a legal benchmark’s partial pass rate.
Bottom line — The largest reported gain was GLM-5’s 11.5-point rise on WorkBuddy-Bench after one refinement step.
Go deeper 5
-
The researchers describe a Propose–Probe–Compose process: propose harness edits, test them on the same task and combine changes supported by observed behaviour.
-
Harness² exposes eight editable components, including guardrails, skills, scripts, subagents and plugins, according to the paper.
-
The framework also retains adopted harnesses and evidence-backed lessons to guide later tasks; each new task starts from the common base harness.
-
The benchmark results are researcher-reported means over three runs, and the maximum gains came from different model and benchmark configurations, according to the paper.
-
The paper reports that parallel recursion with a larger search budget produced further gains, alongside additional computation.