Google Research has released a guide for the RRSI framework, which automates the evolution of AI agent harnesses, including prompts, tools, and memory, around a frozen model. The system uses Claude Opus on Vertex AI to draft edits, which are then validated through Docker-based benchmarks.
The framework evaluates candidates based on two primary metrics: S, the average reward across tasks, and C, the mean token usage per trial. To prevent gaming the system, RRSI treats crashes during evaluation as zero-reward events, ensuring that agents cannot improve their scores by failing on difficult tasks.
RRSI employs a calibration process to determine a 'noise band' (delta), which distinguishes genuine performance gains from random variance. Gains smaller than this threshold are considered statistically insignificant, preventing the system from accepting changes that do not provide a clear improvement.
The selection process uses a strict rule set to admit candidates. It rejects any agent performing below the historical best score minus the noise band. Furthermore, candidates must justify increased token costs with sufficient performance gains, while cheaper, slightly lower-scoring agents may be accepted if they fall within the noise band.
To maintain stability, the framework uses a 'best-ever' score anchor that only increases over time. This prevents the system from drifting toward lower performance through a series of cost-saving swaps. Additionally, an annealing edit budget limits the number of coordinated edits allowed in later stages of the evolution.
