Research on "self-evolving AI agents" — agents that autonomously improve themselves even after training — is being systematized. A taxonomy spanning single-agent optimization, multi-agent collaboration and domain-specific adaptation techniques has been compiled through several academic surveys and an accompanying open-source resource collection, and made available to researchers and developers.
Self-Evolving Agents · EvoAgentX
Mapping the AI Agents That Rewrite Themselves
A newly published taxonomy — "Awesome-Self-Evolving-Agents" — organizes a 2023–2025 research surge into three axes: single-agent optimization, co-evolving multi-agent systems, and domain-specific adaptation. It frames a shift from static, deploy-and-freeze agents toward ones that keep learning on their own.
3
Axes of the taxonomy: single-agent, multi-agent, domain-specific
77
Pages in the flagship survey, classifying hundreds of papers
'23→'25
Research surge feeding the curated repository
Four ways an agent can evolve itself
The survey axes: what , when , how , and where evolution happens.
WHAT
Models, memory, tools, prompts, workflows, agent population
WHEN
During task execution or between runs
HOW
Reinforcement learning, evolutionary algorithms, textual feedback
WHERE
Single-agent, multi-agent, or domain-specific
Static agents vs. self-evolving agents
Conventional approaches are essentially fixed after deployment. Self-evolving agents add a feedback loop that keeps improving.
Static agents
DSPy · prompt tuning
AutoGen · CrewAI (fixed roles)
RAG · fixed structure
Optimized once, then frozen
→
Self-evolving
Evolve parameters
Evolve memory & tools
Evolve workflows
Evolve the population
Continuously improves via feedback
The self-improvement loop
Generate
→
Execute
→
Evaluate
→
Evolve &
Automated in the EvoAgentX framework — demos include a financial analysis assistant and an arXiv paper recommender.
Received favorably
A convenient one-stop list consolidating scattered papers and tools; the surveys are praised as a systematic "roadmap toward ASI."
Open challenges
Still "cooking": real-world deployment safety, immature evaluation metrics, and unstable co-evolution — large-scale evaluation yet to come.
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article…