Google Research releases Harness², code that tunes an agent's harness without retraining
Google Research open-sourced Harness², a framework that refines an agent's prompts, tools and memory per task without ground-truth checks or weight changes.
Google Research has published Harness², an Apache 2.0 framework that automatically edits the system around a model, its prompts, tools, memory and skills, while leaving the weights fixed. The point for builders is that it refines the harness without ground-truth feedback. If you've been hand-tuning an agent's system prompt and tool guidance, this is an early look at automating that loop.
According to the README, each task goes through a Propose–Probe–Compose step: propose edits, probe their effects on the same task, then compose a refined harness from the differences between runs. A second loop keeps adopted harnesses and "evidence-supported lessons" to guide later tasks. Eight components are editable, including guardrails, subagents, scripts and plugins. It runs with OpenCode by default or Codex via --substrate cx.
The paper reports that a single step improved GLM-5 and Gemini 3.5 Flash over their base harnesses, by up to 11.5 points on WorkBuddy-Bench, 4.5 on JobBench and 11.8 percentage points on LAB. Those are research benchmarks, not your app, and we found no independent reproduction.
To try it you need Linux, Python 3.12 and uv. harness2-demo runs locally with no model credentials. Model-backed runs use a .env set up for Vertex AI and incur API costs. The README warns that executor network access is unrestricted and Codex runs with its sandbox disabled, so lock down the network first. It is "not an officially supported Google product."
CCTest's explainer notes the trade-offs: probing edits consumes extra inference, and preferring one behavior over another is not a verified outcome. Whether lessons transfer well across domains is still an open question.