Home~/writing

Tactical: How I think about agents using products and websites

Agents don't guess what they can do. Tell them, in plain text, on the page they land on.

It’s clear that agents will be the major user of the web and they clearly use the web differently than humans. They’re able to consume information faster, more reliably, and take action much faster. This requires a new way to think about how to treat them and to speak their language.

What language they speak

Agents, simply stated, are web native. They speak in plain text (yes, they can consume images/videos, but it’s not as good). They act in HTTP, JSON and APIs. They’re multi-language and are able to read all the content on a site in seconds, while learning how to act.

Agents act on things they’re told to act on

They don’t guess at what you can do – they find what they can do. The principle of: “Tell the agent what it can do” matters the most in determining how agents will use websites going forward. Agents find, they don’t know. When an agent starts to crawl or visit a website, it will nearly always start on the home page – whatever information it was given. This can be from a web search (via a tool like Exa, Tavily) or from a prompt it was given. From there, it does not know where to go. It does not magically know to go to llms.txt, or /.well-known/something-magic, or your markdown version of the page. It needs a signal, a direction, to go and do that. It needs to be explicitly told.

This is where it gets weird and breaks some of the old patterns of the web I knew and have deeply internalized. Most people who know the web, SEO, Google crawling, etc.. know that a crawler would either (a) look at sitemap.xml or (b) follow links on every page and crawl the whole site. Agents don’t do that. They’ll read the home page and depending on their goal, might crawl a different page they find, etc.. but if they don’t find the information they need quickly, they’ll abandon them eventually. If you don’t give them the information they need, they can easily get lost. This is why it’s so critical to include a link to llms.txt directly on the home page – and ideally on every page of your site. You have to put the signpost right where the agent lands. A simple link in the header, footer, or HTML head that tells the agent “Read llms.txt for site documentation and available actions” gives it the exact direction it needs to keep going.

What they actually see when they show up

Even if you put the right links on the page, the agent still has to be able to read them. Most agents today aren’t spinning up a full headless browser to visit your site – it’s too slow, too expensive, and unnecessary if they’re just trying to gather context. Instead, they’re making a simple HTTP request, fetching the raw response, and converting the HTML into markdown or plain text.

If your website or product relies entirely on client-side JavaScript to render content, the agent shows up, sees an empty page shell, and moves on. Server-side rendering (SSR) and clean, readable HTML aren’t just for traditional SEO anymore – they’re table stakes for agents. Better yet, once an agent lands on a page, tell it right at the top that there’s a clean markdown version available (like appending .md to the URL). Give it the raw text it actually wants without forcing it to wade through navigation bars, nested divs, and styling noise.

Moving from reading a website to using a product

Consuming content is only half the equation. The bigger shift happens when an agent moves from reading your website to actually taking action in your product.

Right now, there’s a lot of focus on browser-based agents that click buttons, fill out forms, and navigate UIs just like a human would. That works as a fallback when nothing else exists, but it’s brittle and slow. Remember what language agents speak: HTTP, JSON, and APIs. If an agent wants to query data, book a demo, or configure something in your product, forcing it to click through five screens of UI is the wrong abstraction.

Instead, your website should act as the bridge to your product’s programmatic surface. When an agent reads your site or follows a link to your llms.txt file, you should explicitly tell it how to act:

  • Point it directly to your API documentation and OpenAPI specs.
  • Tell it if you have an MCP server available and how to connect to it.
  • Explain how identity and authentication work (where to get a token, how to pass credentials, etc..).
  • Give it direct endpoints for the core actions a user would normally take in the UI.

When you do this, the website stops being just a marketing brochure for humans and becomes an onboarding flow for agents. The agent lands on the site, reads the instructions, discovers the tools or APIs available, and immediately knows how to use the product.

What this looks like tactically

You don’t need to wait for every new agent protocol or web standard to settle before making your product usable for agents. (The latest, personal agent protocol was announced just last week) Because they just need clear instructions in plain text, you can solve most of this today with a few tactical changes:

  1. Put explicit signposts on every page. Don’t assume agents will guess your URL structure or probe hidden directories. Link directly to your llms.txt, API docs, and agent instructions in your main HTML layout with clear text explaining what those links do.
  2. Serve real HTML first. Make sure your core content and navigation links are rendered on the server. If a basic curl request can’t read your page and see where to go next, neither can most agents.
  3. Offer markdown natively. Expose clean markdown versions of your documentation and core pages, and explicitly tell the agent in the HTML that the markdown version exists.
  4. Map UI workflows to APIs and tools. Audit the main actions a human takes on your site – searching inventory, triggering a workflow, managing an account – and document the direct API or MCP equivalent inside your llms.txt so the agent can skip the UI entirely.

Agents are pragmatic users. They don’t care about your design system, your animations, or your clever copy. They care about finding what they can do and executing on it as fast as possible. If you speak their language and tell them exactly where to go, they’ll do the rest.

← ~/writing