A person opens a docs site in a browser. They scan the sidebar for the section they need, type a word into the search box, skim the headings on a page, and open a related page in a new tab. An agent does none of this. It has no browser. It fetches a file, reads the whole thing, and looks for a direct path to the answer it needs. It also has to cite what it used, which means it needs a stable reference back to where each passage came from.
Your documentation site now serves two readers with different habits. The pages people read are one surface. The files agents read are another set of surfaces, and they all have to say the same thing. When they do not, an agent hands your customer an answer that no longer matches your product.
How an agent reads a docs site
A person navigates by sight. They use the sidebar to find a section, the search box to find a page, and the headings on a page to find the part they want. When one page is not enough, they open related pages and compare them.
An agent works from files rather than from a rendered page. It fetches a text file and reads the whole content at once. It wants structure, so it can find the passage that answers its question without rendering anything. It wants a direct route to the answer, not a browse. It also needs to cite its source, so it needs a stable reference back to the page the passage came from.
The pages you write for people do give an agent something to read. A rendered page has text an agent can parse. That text is wrapped in layout, navigation, and scripts the agent has to work around. The surfaces built for agents strip that away and hand the agent the content in a form it can use directly.
The surfaces agents read
Several plain-text surfaces sit alongside the pages people read. Each one gives an agent a different way in.
The first is llms.txt, a plain-text index of the whole site. It lists where the pages are and what each one covers. An agent reads it to learn the shape of your documentation before it goes deeper.
The second is llms-full.txt, a single plain-text file that contains the whole documentation. Some agents read everything at once rather than fetching pages one at a time, and this file serves them.
The third is a Markdown version of every page. Each page exists as one Markdown file with the content but none of the layout. An agent that wants a specific page fetches its Markdown instead of parsing the rendered HTML.
The fourth is an MCP server. MCP is a structured protocol that lets an agent search the docs and retrieve a passage while it works on a task. The agent does not have to fetch and read whole files. It asks the server for what it needs and gets the matching passage back.
The fifth is WebMCP, a newer standard for agents to interact with a website directly. It is recent, and adoption is still forming. It may or may not last, but it is one of the surfaces agents are starting to expect.
The sixth is a skills index, a list of the tasks the documentation can teach an agent to do. Each entry points at the pages that cover that task. An agent reads the index to find out what it can learn to do on your platform.
| Surface | What it is | What an agent uses it for |
|---|---|---|
| llms.txt | A plain-text index of the site | Learning the shape of the docs before going deeper |
| llms-full.txt | The whole documentation in one file | Reading everything at once |
| Markdown version of every page | One file per page, content only | Fetching a specific page without parsing HTML |
| MCP server | A structured search and retrieval interface | Asking for the passage it needs while working |
| WebMCP | A newer standard for interacting with a site directly | Acting on the site, where supported |
| Skills index | A list of tasks the docs can teach | Finding what it can learn to do on the platform |
Every one of these surfaces comes from the same source files as the pages people read. They are not hand-maintained copies. They are generated from the same content.
- One source
The documentation files describe the current product.
- Human and agent readers
Publish a rendered site and machine-readable surfaces.
- Every change
Regenerate the pages and their agent-facing versions.
- Check consistency
Test that every surface matches its source page.
Why surfaces drift, and what keeps them current
A surface generated once and never regenerated drifts. The page it came from gets edited, and the surface does not. A stale llms.txt can describe a field the product has since renamed. An agent reads the old name, hands it to your customer, and the customer gets an error. The surface now contradicts the page, and the agent has no way to know.
The way to keep them in sync is to regenerate every surface on every documentation change and then test that each surface still matches its page. When a page changes, its Markdown changes, the index changes, and the full-text file changes, all from the same edit. A test compares each generated surface against the source and fails the change when they disagree.
Standards add a second source of drift. WebMCP is recent, and no one knows which of these standards will last. A site that was readable by agents a year ago may not be readable now, because the agents moved to a different surface. The surfaces a docs site offers have to track the standards as they form.
What SaturnDocs does
SaturnDocs is a documentation platform built for the AI agents that now read your docs, and for the people who still do. The agents write, publish, and maintain the documentation site for a software company. On every change they regenerate every surface, from the Markdown pages to the llms.txt index, the full-text file, the MCP server, WebMCP, and the skills index, and they test that each one matches the pages people read. They watch new standards as they appear, test them, and deploy them across every customer site, so a site that is readable today stays readable as the standards change. You supply the product context, and you approve each change before it publishes. If you want to see how your docs hold up for an agent reader, send us a link and we will run them through every surface and send you the report.