A Field Guide to Agents That Fix Themselves

What it does This repository is a curated reading list and visual taxonomy for the flood of papers on self-evolving AI agents. It accompanies an arXiv survey and sorts recent research—mostly 2023 to 2025—into single-agent optimization, multi-agent systems, and domain-specific refinement. Essentially, it is an attempt to impose file-system order on a literature that is expanding faster than most file systems.

The interesting bit The taxonomy deliberately splits “self-evolution” into training-time methods like RL and SFT, test-time strategies like tree-of-thought search, and prompt-level optimization. That distinction matters because the field is currently throwing compute at both model weights and inference-time reasoning without always labeling which is which.

Key highlights

• A tree-structured taxonomy diagram categorizing agent evolution techniques from 2023 through 2025.
• Heavy coverage of test-time compute scaling, from chain-of-thought variants to process reward models.
• Prompt optimization elevated to a first-class track, alongside model training and multi-agent collaboration.
• Links to code repositories provided next to arXiv papers where available.
• References companion open-source frameworks EvoAgentX and MASLab, though their code lives in separate repositories.

Caveats

• The README openly admits the project is “still cooking,” so the bibliography is incomplete and likely favors very recent preprints.
• This repository contains no runnable code or comparative benchmarks; it is strictly a directory of papers and diagrams.

Verdict Bookmark this if you are building agents and need to trace whether a self-improvement trick comes from RL, search, or prompt engineering. Skip it if you are hunting for a drop-in library to evolve your own agent tonight.