Self-improving Agent Harnesses Overfit. How Google Fixes It

Let an agent rewrite its own harness long enough, and it starts looking brilliant on the same tasks it learned from. Then you move it to a benchmark it never saw, and the magic leaks away. That is the pattern Google Cloud AI Research says it found after running four recent harness-evolution methods on one setup: the best method on the evolve set barely improved anywhere else, and two of the four were worse than the harness they started with.

The setup is the thing around the frozen model: prompts, control flow, tools, memory, context handling, the plumbing that decides whether the model can actually do useful work. Methods such as Meta-Harness, AHE, and Self-Harness automate changes to that plumbing. An LLM proposes edits from failed runs, the harness gets scored again, and if the score goes up, the change sticks. The catch is obvious once you say it out loud: the same finite eval set gets reused over and over, so the search learns the test. Google calls that adaptive overfitting, and splits it into benchmark-specific fitting, noise chasing, and complexity accumulation.

Their answer is RRSI, which keeps the edit space open but puts guardrails around how edits survive. On the proposal side, it limits how many independent changes can be bundled together, logs edit history so the search does not keep retrying dead ideas, and forces exploration when progress stalls. On the selection side, a critic screens out benchmark leakage before scoring, the system measures its own eval noise by rerunning the untouched harness, and an edit only sticks if the score clears that noise floor and the added token cost stays within budget. Components that stop paying for themselves get pruned.

The numbers tell the story. The unregularized approach reached 92.8 on the evolve set but only 40.3 on held-out benchmarks, compared with 39.7 for the original harness. RRSI scored a smaller 90.5 on the evolve set but 43.6 held-out, the only method to beat the starting harness by more than a point out of distribution. It also used 2.42 million policy tokens per trial instead of 3.80 million for unregularized evolution.

Google’s ablations sharpen the point. Remove the acceptance rules and the search spends more tokens for less transfer. Remove the proposal rules and the evolve score hardly changes, but held-out performance slips. The same pattern shows up outside judge-graded tasks too: RRSI lifts EngDesign and Frontier-Eng, and the harness transfers across Gemini models as well. The gain is not huge, and the method still costs more than doing nothing, but that is the honest trade: a little less benchmark cosplay, a little more actual generalization.