Google Unveils RRSI Framework for AI Agents to Improve Their Own Harnesses

Google Cloud AI Research has unveiled RRSI (Regularized Recursive Self-Improvement of Agent Harnesses), an open-source framework that lets AI agents fine-tune their own harnesses without modifying the underlying language model. This innovation, introduced in collaboration with UNC-Chapel Hill, Stanford, and Washington University in St. Louis on September 21, 2026, allows agents to propose edits to their harness, test them, and retain the improvements.

RRSI addresses the overfitting problem by constraining edit proposals and selecting only viable changes. The framework employs temporally annealed edit budgets, which grant more room for sweeping changes early on and gradually push toward smaller tweaks as the process matures. History-conditioned exploration is also used to consider previous attempts when proposing new edits.

The selection side of RRSI incorporates four filters: a leakage critic that screens for benchmark-specific knowledge, a noise floor that ignores minor gains, a cost rule that weighs improvements against their costs, and pruning that trims away underperforming components. On Terminal-Bench 2.1, scores climbed from 74.2% to 80.2%, representing a gain of 6.0 points.

RRSI's efficiency is notable, with a reported 30% reduction in policy tokens compared to unregularized evolution methods. The framework also enables transfer across models, allowing harnesses evolved using Gemini 3.5 Flash to improve performance when paired with other model variants.