AI news · October 16, 2026
Google Research open-sources Harness² for recursive agent improvement
Google Research released Harness² under Apache 2.0, with reported gains up to 11.8 percentage points on a lab benchmark, so builders can test recursive agent improvement without retraining model weights.
- reported LAB gain
- 11.8pp
- reported JobBench gain
- 4.5
- license
- Apache 2.0
Google Research released Harness² as an Apache 2.0 repository for improving agents while keeping the underlying model weights fixed. Its task-level loop is Propose, Probe and Compose: an agent suggests a better process, tries it, keeps useful parts and stores the result for later tasks. The project supports parallel and sequential recursion and names OpenCode and Codex executors. Google reports gains of up to 11.5 points on WorkBuddy-Bench, 4.5 points on JobBench and 11.8 percentage points on a LAB partial-pass measure.
This is closer to an experiment manager than a new model. It can make a tool-using system better at a repeatable task, but stored improvements can also preserve bad assumptions or increase call cost. The reported results come from the project, so an independent task set matters. A builder should compare a fixed baseline with the harness on the same inputs, with a hard budget and human review of each accepted process.
What you can do with it
Try the GitHub repository on one repeatable task such as issue triage or document extraction. Freeze the model and tool list, run a baseline, then allow one recursive improvement round. Record accepted-result rate, retries, token cost and the number of manual corrections.
Our take
Harness² is useful if the task repeats enough to repay the extra search. The fixed-weight approach makes the experiment accessible, but reported benchmark gains do not show that every product benefits. Keep the search budget small and reject any process that raises cost without improving accepted work.
Links Try it
Source: Google Research — Made With Models writes the brief; the reporting is theirs.