The Path to Recursive Self-Improving Agents: Foundation, Framework, and Future Directions
Published in Preprints, 2026
As AI agents become increasingly capable, a central question is emerging: can agent systems move beyond manual refinement and autonomously improve themselves? This survey studies self-improving agent systems — systems that transform experience and evaluation feedback into persistent updates to their own components. We formalize the agent system as a coupled evolving state of the foundation model, agent harness, agent data system, agent trainer, and improvement mechanism, and introduce a five-level capability grading standard (L1–L5), ranging from manual improvement to general recursive self-improvement. Building on this foundation, we propose a unified research framework for analyzing self-improving agent systems across their core components, dependencies, and improvement procedures, and organize existing work into a structured taxonomy covering agent harness self-improvement, agent data system self-improvement, agent trainer self-improvement, and cross-component co-improvement. Finally, we discuss key open problems on the path toward recursive self-improving agents, spanning long-horizon evaluation, modifiable infrastructure, generalizability, safety, and human–agent co-improvement. We hope this survey serves as a roadmap for researchers and practitioners exploring the next stage of RSI agent systems. [Project Page] [GitHub]