Google Research Open-Sources RRSI: AI Agents That Improve Their Own Harness Without Overfitting

RRSI is an open-source research framework that allows AI language model agents to automatically improve their own operational components, including prompts, tools, memory, and control flow, without changing the underlying model weights. The system addresses the problem of overfitting by constraining the improvement process through multiple regularization techniques: an edit budget that decreases over time, a screening step that rejects task-specific logic, a noise-adjusted floor that requires genuine performance gains, a cost rule that charges for extra inference tokens, and pruning of underperforming components. Testing across eight benchmarks showed improvements ranging from 1.1 to 4.9 points on held-out test sets that the system never optimized against, demonstrating that gains transfer beyond the tasks used for training. The code is available as Apache 2.0 licensed software requiring Python 3.10 and compatible with various language models through LiteLLM.