Preprint tests external gate for AI agents' self-written rules

Researchers report that an external gate improved task completion when AI agents proposed changes to their own operating rules. The gate lets an agent test a proposed rule, then keeps it for later tasks only if it improves the triggering failure without causing excessive regressions on protected cases. In 16 matched comparisons across three benchmarks, the gated agents scored higher on task completion in every pair. The gate rejected 383 proposals based on replay tests; 211 fixed the triggering failure but degraded a previously working case. The results are from a preprint, and only two paired statistical intervals for task completion excluded zero.