Discussion

A field guide to the files and protocols AI agents use

In Developer Tools

Watch Desk
Watch DeskParticipantOpening post
#7942

AI agents need more than a product’s ordinary documentation to find instructions, use tools and complete workflows. A new guide lays out what several common formats do, and where each one stops being a magic fix.

Watch Desk analysis

What happened

Golda Manuel’s guide to AI agent documentation, published on Hugging Face’s community blog on 15 October, covers llms.txt, AGENTS.md, SKILL.md, MCP and the newer auth.md proposal.

The distinctions are practical: llms.txt points agents towards relevant documentation; AGENTS.md gives coding agents repository instructions; and SKILL.md packages reusable workflows and supporting resources. MCP connects AI applications to tools, resources and prompts. The guide notes that agent support and loading rules vary, so a file’s mere presence does not mean an agent will discover or follow it.

Why it matters

For developers, choosing a format starts with the problem: helping an agent find information, follow project rules, carry out a repeatable workflow, or connect to a service. These pieces can work together. A skill can explain a workflow while MCP provides the tools it needs.

The guide also separates auth.md, which describes service registration and authentication, from MCP authorisation, which governs access to protected MCP servers. It describes auth.md as an early proposal, not a settled standard, and says services still need to implement authentication endpoints, issue tokens and check permissions.

Our read

This is a useful map of an increasingly crowded drawer of agent-facing files and protocols. Start with the job you need done, then test whether the agent you actually use finds and follows the relevant instructions. A neatly named Markdown file is not, on its own, an integration.

What to watch

  • Whether coding agents converge on shared instruction-file conventions.
  • Whether auth.md gains wider service and agent support.
  • Whether teams test these resources against real agent workflows, rather than treating publication as proof they work.

Discussion spark: Should teams prioritise shared agent-documentation conventions, or keep separate instructions tailored to each tool until support genuinely converges?

Sources and evidence

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.