Google's RRSI lifts unseen-task scores by up to 4.7 points

The research paper argues that much of the recent progress in AI agents comes from work on the harness — the scaffolding that decides whether an agent reads the right file before changing it, recovers from mistakes, and delivers results cleanly — rather than from new models. Until recently, improving this harness was done by hand: people reviewed failed runs and patched it manually. Newer methods automate the loop by having a language model rewrite the harness itself, based on feedback from the test tasks, which the researchers call a practical form of recursive self-improvement.

The paper shows this self-optimization comes with a catch. Because the agent keeps working on the same limited set of test tasks, it ends up memorizing them: scores on training tasks go up, while gains on new, unseen tasks shrink or disappear. The researchers say this happens because the search memorizes patterns that only fit one benchmark, favors candidates that score well purely by chance, and piles on unnecessary complexity that raises the test score without making the agent any better. RRSI addresses this at both ends of the optimization loop — when proposing new changes and when deciding which become permanent — by shrinking the edit budget over time and having a critic throw out any proposal that hardcodes task names, solutions, or other benchmark-specific tricks.

Among the optimized harnesses, RRSI needs the fewest tokens and steps and performs best on new tasks, though the unmodified baseline harness is even leaner. The authors note their study only covers harnesses built around frozen models and doesn't address cases where the model weights change, so whether these guardrails help when the model itself is being updated remains an open question.