Google paper: auto-optimized agent harnesses overfit evals — RRSI fixes it

omarsar0 · x · 2026-09-24

A new Google paper shows that if you auto-optimize an agent's harness (prompts, control flow, tools, memory), eval scores can go up while the agent gets worse on real tasks.

The problem: automated harness evolution proposes edits, keeps the ones that raise the score, and repeats — effectively overfitting the training tasks. Of five harness-evolution methods compared, Meta-Harness hit 93.0 on the Harvey LAB evolve split but gained only 0.3–1.5 points on JobBench, GDPval, and APEX-Agents.

The fix (RRSI): adds regularization on both sides of the loop — the proposer gets a shrinking edit budget and is pushed toward untried directions, while a critic rejects benchmark-shaped edits.

Result: RRSI scores lowest on the tasks it evolved against but highest on all three out-of-distribution benchmarks, showing how to prevent harness self-evolution from overfitting evals.

Related event: Google's RRSI Regularizes Recursive Self-Improvement of Agent Harnesses(8 posts)→

Original post →

More from coding & agent

coding & agent channel →