Google releases Harness² for runtime improvement of AI agent workflows
The AMW Read
Reusable harness-improvement experience meaningfully extends agent orchestration, with reported gains across models and tasks but unresolved production reliability and operating costs.
Google releases Harness² for runtime improvement of AI agent workflows
Google has open-sourced Harness², a system that revises an agent's supporting code, prompts, tools, and skills during task execution, according to QbitAI. It supports OpenCode and Codex. Its task-level loop proposes general and domain-specific changes, probes candidate harnesses against a reference execution, and composes changes using behavioral evidence. A second loop records lessons in an improvement library. Subsequent tasks start with a clean base harness while the improvement process retrieves relevant experience.
The reported results make agent orchestration a concrete competitive variable alongside model capability. On WorkBuddy-Bench, one improvement step raised GLM-5's OpenCode score from 64.2 to 75.7, above the unmodified Gemini 3.5 Flash score of 72.4. Gemini also improved, reaching 77.5 from 72.4 with OpenCode. These comparisons suggest that execution design can change model rankings within a benchmark. They support the case for value in orchestration and reusable operational experience, but do not establish reliable autonomous improvement in production: comparing execution behavior without a ground-truth answer leaves correctness uncertain.
For builders, the concrete implication is to evaluate the model and harness together. Harness² separates temporary task repairs from accumulated improvement knowledge, addressing the risk that one task's modifications impair the next. Teams testing it should measure verified output quality, latency, and total inference cost together. The source reports additional gains from five parallel candidates, but provides no corresponding cost or latency figures, leaving the economic value of that extra exploration unresolved.
