Recursive Self-Evolving Agents Improve LLM Performance Safely

Improving LLM agents often involves evolving natural-language artifacts like prompts or workflows without updating model weights. This study introduces RSEA (Recursive Self-Evolving Agent), which maintains a three-layer natural-language state: strategy, skills, and a playbook. RSEA iteratively rewrites these layers based on its own trajectories, but critically, only commits changes if they do not degrade performance on a separate held-out dataset. This "keep-better" gate ensures safe, monotonic improvement. The research evaluated RSEA against several baselines across four benchmarks, demonstrating that while no single artifact universally wins, RSEA consistently performs well and, unlike some unguarded methods, avoids catastrophic performance collapses. Its held-out selection mechanism is key to its safety and reliability.