📄New Research on Self-Evolving Agents: When AI agents modify themselves, how…

An agent's self-modification needs evidence that the modification cannot rewrite. That is the central reliability principle in a self-evolving-agent survey whose authors include researchers at Tencent. Its framework separates producing a change from deciding whether it should persist: an external audit can accept it, reject it and roll it back, or escalate it for human review. Survey project.

The distinction matters most when an agent can modify its own evaluation criteria, because a better score could then reflect a changed test. The survey's five levels, L0 to L4, classify how deep an update reaches, from revising a single output to changing how updates are judged; they are not a ranking of intelligence and do not establish reliability. Taxonomy and reliability framework.