The Pulse
Google Research Releases Harness² for Agents That Rewrite Their Setup
Google Research released Harness², an open-source framework that lets AI agents test changes to their prompts, tools and other operating components while keeping model weights fixed. Its paper reports gains across three professional benchma
AI.info Team · Written with AI ·
LAB’s partial pass rate rose 11.8 percentage points
Harness² improved results by as much as 11.8 percentage points on LAB, a legal-work benchmark, according to the researchers behind the project. The figure measures LAB’s partial pass rate, not a general increase in an agent’s accuracy. The paper also reports gains of up to 11.5 points on WorkBuddy-Bench and 4.5 points on JobBench.
Google Research made the framework’s code available in a public repository. Its accompanying paper, Harness²: Recursive Agent Harnessing for an Open World, is dated October 9, 2026. The system changes the software around a model rather than updating the model itself.
Harness² edits eight parts around a fixed model
An agent harness is the surrounding system that shapes how a model acts: its prompts, tools, memory, skills and orchestration, among other components. Harness² exposes eight editable parts, including guardrails, scripts, subagents and plugins. Its experiments use GLM-5 and Gemini 3.5 Flash with their weights held fixed.
The framework runs a task-level loop called Propose–Probe–Compose. It proposes edits, tests them on the same task and combines changes supported by contrasts between runs, without relying on a ground-truth answer key. A second loop saves adopted harnesses and lessons backed by those tests, so later tasks can draw on earlier refinement.
“We introduce Harness^2, a plug-and-play, recursive agent harnessing framework that addresses these challenges through two connected levels of recursion.”
Ruiyao Xu et al., Harness²: Recursive Agent Harnessing for an Open World
Users can run recursion in parallel, exploring several candidates, or sequentially, carrying one revision forward through successive steps. The repository supports OpenCode and Codex executors and includes workflows for JobBench, LAB and WorkBuddy-Bench, which covers office, web and coding tasks.
Benchmark gains do not guarantee a better run
The paper reports that one improvement step beat the starting harness on all three benchmarks across four model-and-harness configurations. More recursion sometimes added further gains, but not every added step improved every result. On WorkBuddy-Bench, for example, the reported extra recursion raised some scores while lowering others.
The authors also document cases where the system selected a weaker candidate or made an edit that narrowed task coverage. In one WorkBuddy-Bench coding example, selection chose an output scoring 55 over one scoring 95; on LAB, a narrower result fell from 54 to 26. Those cases show why the system’s own execution evidence is not the same as a reliable correctness check.
The paper says the feedback process depends on behavioral contrasts and model judgment, which can be affected by execution variability and shared model biases. The authors caution that an adopted change is not guaranteed to be correct or to transfer to a different task. Their evaluations cover three benchmarks, not every kind of work agents may encounter.
Model-backed runs add cost and network exposure
The repository offers a local demo using deterministic editors and Python executions without model credentials. Running model-backed benchmarks requires additional setup, and the README warns that these runs incur API costs. The project is licensed under Apache 2.0, while third-party code, benchmarks and datasets retain their own licenses.
There is also an evaluation risk: the README says executor network access is unrestricted and Codex runs with its sandbox disabled. Public benchmark rubrics may be reachable online, so the project tells users to configure network restrictions before evaluation to keep agents from accessing answers or grading criteria.
Google’s repository also states that Harness² is not an officially supported Google product. For teams testing the approach, the immediate question is whether its extra model calls and harness edits produce improvements that survive tasks beyond the benchmark runs.