Self-Evolving AI Agents (survey)

“A Comprehensive Survey of Self-Evolving AI Agents: A New Paradigm Bridging Foundation Models and Lifelong Agentic Systems” — Jinyuan Fang et al. (arXiv:2508.07407). The academic framing behind this spoke’s self-improving-agents node, which had rested on vendor READMEs.

Definition and framing

Self-evolving agents “automatically enhance their capabilities through interaction data and environmental feedback, rather than relying on static manually-crafted configurations.” The survey positions this as bridging “the static capabilities of foundation models with the continuous adaptability required by lifelong agentic systems” — a research-grade statement of the “growth axis” the wiki tracks.

The unified framework

The survey organizes the field around a feedback loop with four components: System Inputs, the Agent System, the Environment, and Optimizers (the mechanisms driving evolution). Different self-evolving techniques target different components of the agent system — so “self-improvement” is not one move but a family scoped by what gets evolved.

Domains and risks

It reviews domain-specific strategies in biomedicine, programming, and finance, and treats evaluation, safety, and ethical considerations as central to “ensuring their effectiveness and reliability.” That names the wiki’s open drift/quality risk as a recognised research problem, not a hunch. An accompanying repo, Awesome-Self-Evolving-Agents, catalogs the techniques.

Tier T1 — peer-track survey (arXiv, CC-BY-4.0). Cited by self-improving-agents, loop-engineering.