A field guide to agents that upgrade themselves

What it does This repository is a curated awesome-list and companion to the survey paper Self-Improvements in Modern Agentic Systems: A Survey (arXiv:2607.13104). It collects papers, benchmarks, blogs, videos, and other media on foundation-model agents that can update themselves without human intervention. The list is organized around a formal taxonomy that classifies improvement mechanisms by whether they target the model’s parameters or its surrounding scaffolding.

The interesting bit The taxonomy deliberately splits “self-improvement” into two distinct camps: Foundation Model Improvement (retraining weights from generated demonstrations or environmental feedback) and Scaffolding Improvement (rewriting prompts, evolving tools, or refactoring memory and workflows). That framing treats prompt optimization and tool synthesis as first-class improvement loops, not just post-training cleanup.

Key highlights

• Accompanies the survey paper with LaTeX source and an agent-oriented reading guide (README_AGENT.md)
• Taxonomy spans intrinsic generative demonstrations, evaluative feedback, and extrinsic exploratory experience for model-level updates
• Scaffold-level categories cover prompt optimization, memory, tool, and full-scaffolding improvement
• Curated tables link directly to papers and code repositories where available
• Includes a timeline figure tracing the evolution of representative self-improving agent systems

Verdict Researchers and engineers who need a structured map of how agentic systems can self-modify will find this a useful on-ramp. If you are looking for a drop-in self-improvement framework rather than a literature index, look elsewhere.