1 min read

Filesystems Are Having a Moment: Why AI Agents Are Rediscovering Files for Persistent Context

Canopy AI

Developer Madalitso Mumba published an essay arguing that AI agents are driving a renaissance in filesystem use as a persistence and context layer, with CLAUDE.md, SKILL.md, and aboutme.md formats emerging as portable, agent-readable alternatives to proprietary SaaS data silos. The essay synthesizes research including an ETH Zürich study showing repository-level context files can increase inference cost 20%+ while reducing task success rates if over-specified, and covers the broader architectural shift Jerry Liu of LlamaIndex identified: toward agents with a filesystem and 5-10 tools over agents with 100+ MCP tools. Anthropic's Agent Skills standard — already adopted by Microsoft, OpenAI, Atlassian, GitHub, and Cursor — is framed as the first serious attempt to make skill files portable across competing AI coding tools.

Key Takeaways

  • ETH Zürich study found repository context files (CLAUDE.md etc.) increased inference cost 20%+ while reducing task success rates — conclusion: context files should describe only minimal constraints, not 2,000-word onboarding docs
  • Anthropic's SKILL.md Agent Skills standard adopted by Microsoft, OpenAI, Atlassian, GitHub, and Cursor — enabling skills written for Claude Code to work in Codex and Copilot without coordination
  • Jerry Liu (LlamaIndex) and Andrej Karpathy both argue the ideal agent architecture is filesystem + 5-10 tools, not 100+ MCP tools — Claude Code's CLI-on-localhost model is cited as the dominant current AI use case

Original source: madalitso.me / Hacker News