Abstract
Agent systems based on foundation models have advanced rapidly and can now accomplish increasingly complex tasks with greater autonomy. However, manually developing and refining agent systems requires substantial human effort and time, motivating a shift toward agent systems that can improve themselves. In this survey, we study self-improving agent systems that autonomously transform experience and evaluation feedback into persistent updates to their components. To compare the self-improvement capabilities of diverse agent systems, we introduce a five-level grading standard, ranging from manual improvement to general recursive self-improvement. Furthermore, we propose a unified research framework that jointly models the foundation model, agent harness, agent data system, agent trainer, and the mechanism governing their improvement as a coupled system. Guided by this framework, we systematically review the literature on self-improvement and provide a taxonomy covering self-improvement in individual components and co-improvement across components. In addition, this survey identifies open research problems on the path toward recursive self-improving agents. Overall, this survey provides a foundation for studying self-improving agent systems and a roadmap for future research.