AI can now evolve its own debugging code — and does it better than humans under tight evaluation budgets

Researchers built a system (RoboPhD) that evolves AI agents by having them compete against each other using only training data, no separate validation step, and it outperforms two other approaches on most benchmarks. This matters because expensive evaluations—human feedback, multiple model calls—become the bottleneck in scaling AI improvement, and this system cuts that bottleneck by using competition itself as the evaluation mechanism.

Every AI system being deployed today started as a prompt or architecture someone wrote by hand, then improved through human feedback or trial-and-error. That feedback loop is slow and expensive. RoboPhD shows that you can skip the feedback step entirely: let agents evolve by competing, and they'll build increasingly sophisticated diagnostic code to help themselves and their successors. The practical implication is blunt: if this holds, the cost of improving AI systems drops, which means more experiments happen faster, which means the systems that get built are determined less by who has the biggest evaluation budget and more by who has the fastest iteration loop.