Description
The paper from a Google team looks at methods that automatically evolve an LLM agent's harness, meaning the prompts, control flow, tools, memory and context management around a frozen model, and shows that this recursive self-improvement tends to overfit the tasks it trains on. RRSI adds regularization to the loop: the proposer works with an annealed budget on how many edits a candidate can bundle and favours unexplored directions, while a critic screens benchmark-specific proposals and a pruner removes changes that are too small, too costly or no longer useful. Across eight coding, agentic workspace and engineering design benchmarks it gains up to 14.1 points in-distribution and up to 4.7 points out-of-distribution, with a harness that uses 30% fewer policy tokens. Useful reading for anyone tuning agent scaffolding automatically.
