A field guide to agents that upgrade themselves

What it does

It is a curated survey that maps research on self-evolving agents onto a structured taxonomy. The README organizes papers by what evolves—models, context, tools, or architecture—when evolution happens, and which mechanisms drive it, from reward-based feedback to population-level selection. Each entry is a link to an arXiv paper, making the repo a navigable index rather than a codebase.

The interesting bit

The taxonomy treats agent evolution as a design space with orthogonal dimensions, distinguishing between intra-test-time adaptation (reasoning harder mid-task) and inter-test-time learning (accumulating skills across sessions). That separation is a useful lens for a field that often conflates quick thinking with genuine learning.

Key highlights

• Covers four evolutionary targets: models, memory and prompts, tools, and multi-agent architecture
• Separates “when” into within-task refinement and across-task lifelong learning
• Surveys mechanisms including reinforcement learning, imitation, and evolutionary search
• Evaluation and future directions sections are present in the outline
• Over 1,200 stars suggest the topic is resonating

Caveats

• The README is largely a table of contents with arXiv links; little original synthesis text is visible
• The only recent update redirects readers to a separate repository on “reasoning economy”
• Several evaluation subsections are commented out in the source, suggesting the document may still be incomplete

Verdict

Worth bookmarking if you are researching agentic AI or writing a literature review; skip it if you are looking for runnable code or frameworks.