AI Agent Documentation: A Guide to llms.txt, AGENTS.md, SKILL.md, and MCP

Documentation, READMEs, and API references have long helped developers get started with new tools. As coding agents take on more of the work, new formats are emerging to help them navigate that information and use developer products successfully.

You've probably encountered some already: llms.txt for finding documentation, AGENTS.md for repository instructions, SKILL.md for reusable workflows, and the Model Context Protocol (MCP) for connecting agents to tools and data. Newer proposals, including auth.md, are beginning to tackle how agents sign up for services and manage credentials. Understanding what each does helps you decide which ones would be useful for your product.

Types of AI agent documentation and protocols

Helping coding agents use a developer product involves more than writing good documentation. Agents need to know where to find information, how to work within a repository, and which tools are available to them. A growing collection of resources aims to make that easier, from documentation files and agent-specific instructions to protocols like MCP. The table below offers a look at what's available, followed by a closer examination of how each works and when it might be worth using.

What is llms.txt and how do AI agents use it?

Finding the right documentation isn't always straightforward for an AI agent. llms.txt is a Markdown file that introduces a product and links to its most useful documentation, giving agents a starting point without having to search through an entire site.

For example, the Quantiles website itself provides an llms.txt that begins like this:

# Quantiles

> Quantiles helps developers evaluate how reliably coding agents complete developer tasks.

## Getting started
- [Quickstart](https://quantiles.io/documentation/quickstart): Get started with agent evaluations.
- [Agent Quickstart](https://quantiles.io/documentation/agent-quickstart): Coding agent instructions to configure, execute, and inspect evaluations.

Agents can follow the links relevant to their task rather than loading everything into context. The proposal also supports discovering these resources through HTML links or HTTP headers and recommends providing Markdown versions of documentation pages. A related convention, llms-full.txt, combines documentation into a single file. Though it seems convenient, a full export may consume unnecessary context. Platforms like Mintlify can generate both files.

Having an llms.txt or llms-full.txt doesn't guarantee that an agent will use it. Evaluate whether agents discover these resources, follow the relevant links, and find the information needed to complete their tasks.

How AGENTS.md guides coding agents

When an agent starts working in a repository, it needs to understand how the project works, from installing dependencies and running tests to following coding conventions. The AGENTS.md format provides a dedicated place for these instructions.

For example, here is the beginning of the Quantiles AGENTS.md:

# Quantiles Agent Guide

This file is for coding assistants such as Codex, Claude Code, Cursor, GitHub Copilot, Gemini CLI, OpenCode, and similar tools. It gives agents a short, public-safe map of the Quantiles repository and the rules for making focused, reviewable changes.

Before attempting tasks in this repository, read these files:

- README.md: public product overview, quickstart, CLI examples, SDK summary, docs links, and agent guidance.
- CONTRIBUTING.md: contributor expectations, development workflow, and review norms.

The instructions depend on the project. They might cover development workflows, testing requirements, or instructions on how to use a particular tool. Unlike website documentation, AGENTS.md is intended to guide agents working in a repository, and not every agent will automatically read it.

Codex layers repository guidance, while other coding agents may load and prioritize them differently. Regardless of the agent, keeping instructions focused and up to date is important. Outdated commands and unnecessary rules can introduce workflow inefficiencies, increase task complexity, and lead to failures when agents interact with your product.

How SKILL.md helps agents complete workflows

Some tasks require more than repository instructions. An agent might need to install a CLI, configure credentials, complete several steps, and verify the results. An agent skill provides reusable instructions and supporting resources to help agents complete these workflows.

quantiles-skill/
├── SKILL.md
├── references/
│   ├── configuration.md
│   └── troubleshooting.md
└── scripts/
    └── verify-install.sh

The SKILL.md file contains a name, description, and instructions for completing a workflow. Supporting files can provide additional documentation or scripts when needed. The description is especially important because it helps agents determine when to use the skill.

Both OpenAI Codex and Anthropic's skills implementation use a process called progressive disclosure to discover and use skills. Agents initially see a brief description of available skills and load the full instructions and supporting resources when needed. For this to work, the skill must be installed where the agent can discover it. Simply adding a SKILL.md file to a repository doesn't mean an agent will find or use it. Even when a skill is available, there's still the question of whether an agent will recognize when to use it and follow the instructions correctly.

How MCP connects AI agents to tools and data

While documentation helps agents understand a product, the Model Context Protocol (MCP) provides a standardized way for AI applications to connect to its tools and data. MCP servers can expose three types of capabilities:

• Tools: Operations agents can invoke, such as creating a project or starting an evaluation.
• Resources: Information applications can retrieve, such as evaluation results or execution logs.
• Prompts: Reusable templates that guide specific tasks or interactions.

For example, a Quantiles MCP server could allow an agent to create and compare the results of evaluations using structured tool calls rather than manually composing API requests or using a CLI. Which capabilities an agent can actually use depends on what its MCP client supports and makes available.

MCPs and SKILL.md files can also work together. A skill can guide an agent through a workflow, explaining which operations to use, in what order, and how to verify the results. MCP provides access to the tools needed to carry out those steps.

Instruction files for different coding agents

The resources we've covered so far, including llms.txt, AGENTS.md, SKILL.md, and MCP, are designed to work across different AI agents, although support varies. But coding agents also have their own instruction files and configuration options. These can differ in how instructions are loaded, where they apply, and which instructions take priority. Here's how some popular coding agents handle them:

• Claude Code: Uses CLAUDE.md for project instructions and .claude/rules/ for more specific rules. It can also support AGENTS.md, depending on the version and configuration.
• Gemini CLI: Uses GEMINI.md for project context and instructions, with options to configure alternative filenames.
• Cursor: Supports AGENTS.md and rules stored in .cursor/rules/. Rules can apply automatically, to specific files, when relevant to a task, or when explicitly requested.
• GitHub Copilot: Uses .github/copilot-instructions.md for repository-wide guidance and .github/instructions/*.instructions.md for instructions that apply to specific paths.

With several agents to support, maintaining separate instruction files can quickly become difficult. Where possible, keep shared guidance in one place and add agent-specific instructions only when needed. And just as with AGENTS.md or SKILL.md, test whether each agent discovers and follows the instructions as intended.

How auth.md supports AI agent authentication

One of the newer proposals is auth.md, introduced by WorkOS in May 2026. Unlike the more established resources we've covered, it's still in the early stages of adoption. It addresses a familiar problem where an agent can find the documentation and understand an API but still get stuck when it needs to create an account or obtain credentials.

The proposal gives services a way to describe registration, authentication, and available permissions in a Markdown file. It supports two main approaches:

• Agent verified: A trusted agent provider verifies the user's identity, allowing the service to register the agent without additional user interaction.
• User claimed: The agent starts registration, then asks the user to sign in and confirm a code to authorize access.

Both approaches ultimately allow the agent to obtain a scoped access token. But publishing an auth.md alone isn't enough. The service still needs to implement the authentication endpoints, token issuance, and permission checks.

It's also worth distinguishing auth.md from MCP authorization. While MCP authorization governs access to protected MCP servers, auth.md focuses on helping agents register with a service in the first place, so that they have an identity that can be later authorized.

As an emerging proposal, auth.md is worth watching, but compatibility and agent support still need to be tested. And even with automated registration, some workflows may still require human approval, particularly when account ownership, permissions, or billing are involved.

Which AI agent documentation standards to use, and how to evaluate them

The best place to start is with the problems agents encounter when using your product. A simple library may only need accessible documentation, while a developer platform with more complex workflows might benefit from agent skills, structured tools, and authentication guidance.

Different resources address different parts of the developer experience:

These resources aren't mutually exclusive, and not every product needs all of them. Adding more instructions isn't always the best solution, either. What matters is whether the ones you do have actually help agents complete developer tasks.

To find out, start by evaluating representative workflows, such as installing your product, completing the quickstart, configuring authentication, or using a more advanced capability. Compare performance with and without the resource to understand where it helps, where it doesn't, and what might need improvement.

Here are a few key questions to consider:

• Discovery and activation: Did the agent find and use the relevant documentation, skill, or tool?
• Task success and reliability: Did it complete the task correctly and consistently across repeated trials?
• Efficiency: How much time, token usage, and tool activity did the task require?
• Errors and recovery: Where did the agent get stuck, and could it recover without human intervention?

As AI agents evolve, documentation will too

AI agents are evolving quickly, and the resources that help them use your product will need to keep up. What works for one agent today may not work as well for another, or even for the same agent after an update.

Keeping a representative set of developer tasks makes it easier to spot those changes. Rerun evaluations as your product, documentation, and agents evolve, and use execution traces to understand where things go wrong. Solving these issues will sometimes require instructions. Other times, it might be a clearer error message, a simpler setup process, or a more useful tool response.

Supporting more formats only matters if they improve the developer experience. What matters most is whether agents can navigate your product and complete tasks reliably and successfully.

Keep reading

View all

Using Token Data to Improve the AI Agent Experience

Evaluations·October 8, 2026

Does your SKILL.md help coding agents?

Evaluations·October 1, 2026

Evaluating AI Agent Tool Use

Evaluations·September 24, 2026