planning-with-files

Adoption judgment: worth watching and testing, but better treated as an agent workflow technique than mature infrastructure. The heat proof is strong: the GitHub repo shows major star growth, and it spread on Zhihu and X. Still, attention does not equal production readiness. Usability proof mainly comes from the official repository, code-level explainers, and a small number of hands-on tutorials, so the evidence base is still limited.

What it does: it stores an agent's plan, discoveries, and progress in files such as PLAN.md, DISCOVERY.md, and PROGRESS.md. That gives Claude Code-like coding agents a persistent state they can reload after session interruption, /clear, or long-running work, reducing repeated trial and error. A better analogy is a file-based memory and execution discipline for agents—not a full automation platform, not a generic vector database or RAG system, and not a standalone multi-agent framework.

Barrier and cost: the project itself is open source in the official repository, so access and usage appear free at the repo level; however, it does not remove model usage costs.