An LLM agent's real-world capability depends heavily on its harness — the prompts, control flow, tooling, memory, and context management wrapped around a frozen backbone model. Recent work has tried to automate harness design by repeatedly proposing and selecting component-wise edits, effectively creating a form of recursive self-improvement (RSI) at the agent-system level.
But that recursion has a failure mode: the harness can overfit by memorizing the training tasks, producing large in-distribution gains that shrink or disappear on out-of-distribution benchmarks. A new paper from Google Research, RRSI: Regularized Recursive Self-Improvement of Agent Harnesses, addresses this by building regularization directly into the evolution process.
RRSI constrains both the proposal and selection stages. The proposer uses a temporally annealed budget that limits how many edits a candidate can bundle, and it encourages unexplored trajectories based on evolution history. The selector adds a critic that screens benchmark-specific proposals and a pruner that removes changes deemed too small, too expensive, or no longer useful.
Across eight benchmarks covering coding, agentic workspace, and engineering design tasks, RRSI gains up to 14.1 points on the split it evolves against and up to 4.7 points on five out-of-distribution benchmarks. The resulting harness also runs on 30% fewer policy tokens than unregularized evolution.
Code is available at github.com/google-research/rrsi, with a project page at regularized-rsi.com.