A new technique curbs how badly AI agents cheat when they redesign their own toolkits.
An AI agent's real capability often comes less from the frozen language model at its core and more from its harness: the prompts, tools, control flow, and memory scaffolding wrapped around it. Some newer systems now edit that harness themselves, proposing and testing changes in a loop, a form of recursive self-improvement applied to the agent's operating instructions rather than the model weights. The problem is that these systems tend to memorize their training benchmarks, posting big score jumps on familiar tasks that shrink or disappear on anything new. Regularized Recursive Self-Improvement (RRSI), described in a paper whose code sits in Google Research's GitHub org, adds constraints to that loop. It limits how many edits a proposed harness can bundle at once, favors unexplored changes over repeats, and prunes edits that are too small, too costly, or clearly tailored to one benchmark.
Tested across eight benchmarks spanning coding, agentic workspace tasks, and engineering design, RRSI gained up to 14.1 points on the benchmark split it trained against but only up to 4.7 points on five benchmarks it never saw during training. That is still a real drop-off, not a fix, but it is a narrower one than unconstrained harness evolution leaves behind, and the resulting harness ran on 30% fewer tokens.
In other words, RRSI curbs the cheating rather than eliminating it, a useful reminder that letting an agent rewrite its own instructions is still closer to a constrained experiment than a self-improving free-for-all.