Google Research open-sources Harness2 to refine agent harnesses after deployment

What the framework changes

A harness covers everything around the model, including prompts, tools, memory, skills and orchestration. Harness2 leaves the model itself untouched and works on that surrounding system. The research team presents it as a plug-and-play framework that can be added to existing agents.

Harness2 exposes eight editable components, each mounted as a file: the system prompt, guardrails, memory, tool guidance, subagents, skills, scripts and plugins. Together they govern context construction, the tool interface, delegation and behaviour inside the tool loop. The layout extends the seven-file split from earlier Agentic Harness Engineering work by one component.

Recursion without labels or tests

The research team argues that most self-improvement methods depend on labels or test cases. Harness2 is built for an open setting in which tasks arrive unseen and no verifier judges whether an answer is right.

The framework runs two linked levels of recursion. At task level, a contrastive Propose, Probe, Compose step first runs the task with the base harness and then evaluates two candidates on the same task: an invariant candidate aimed at general failures such as planning or tool use, and a domain candidate aimed at knowledge and procedures for the current domain.

Reported results and requirements

According to the report, one refinement step improved all four model and harness setups on all three benchmarks. The weakest result was GLM-5 under Codex, which gained only 1.2 points on JobBench, rising from 23.6 to 24.8. The framework trains no parameters.

It runs on Linux with Python 3.12 or 3.13 and uv, and model calls go through Vertex AI.

Open questions

• Do the gains hold on benchmarks beyond the three reported?