Awesome Self-Evolving AI Agents
An open survey and evidence map for judging whether an AI agent system genuinely improves itself from feedback, rather than only appearing smarter in a demo.
This repository is a survey of research and practice on self-evolving AI agents: systems that are supposed to change their own prompts, memory, workflow, code, or weights based on feedback. Instead of only listing links, it starts from a judgment framework asking what changed, why, who verified it, whether the change is retained, and whether it can be rolled back.
It organizes the field into five categories of evolution loop, specification-to-execution, search, evaluators, reflection and memory, and population or archive methods, and maintains a benchmark matrix for code self-improvement systems alongside project write-ups that look at each one's evaluator, archive, and lineage. An accompanying index is described as a working, editorial prototype for weighing evidence of benchmarks, closed loops, and governance, not a settled score. Most long-form material is Chinese-first, with an English section covering the core definitions and framework.