What Can AI Agents Do on Your Website?

You ask an AI agent to do something it did yesterday. It starts a new session, looks around, and has no idea what happened last time.

That can feel a bit like working with someone who forgets every conversation overnight. You can explain the same thing again, or leave a note where the next session will find it.

Hi! I'm Bob Michaels, and this is Evolving the Web. Every two weeks I talk through the essays I just published, plainly, as one story. This time, I'm connecting four essays about what happens when software has to remember, measure, and make its way through the web.

I run many client projects at the same time, with AI agents doing much of the work under my direction. I built a system for this called HALO. The basic challenge is that an agent starts each session with no memory of the last one. Anything it needs to know has to be in a file it reads at the start.

So when a project makes a mistake, I write down the lesson once: what went wrong, the fix, and where it happened. I leave out the client's name so the lesson can help elsewhere. The next session in every other project gets that lesson automatically.

That changes how the work begins. Before an agent starts, its plan has to say how each lesson applies. If it skips one, a check refuses the plan. I like the lesson about proving a check works by watching it fail on purpose first. It's a little like testing a smoke alarm before you need it, rather than waiting for burnt toast to run the experiment.

That same question, what can software find and remember, shows up on public websites too. A plain-text file called llms.txt can sit at the top of a site and tell AI models what the site is and where its important pages are.

A study in August found about one in eleven of the biggest websites have one. Google's search team says Google Search ignores these files, while Google's own Lighthouse audit tool checks whether a site has one. That's a mixed signal from the same company, and it leaves business owners wondering which way to go.

My take is simple: the file takes an afternoon to write, so write it. But it isn't what gets a business recommended. The pages that answer a buyer's questions are what matter. Think of llms.txt as a note pointing to the shop's shelves. The note can help someone find the right aisle, but it can't stock the shelf for you.

And if you want an AI system to do more than find a page, it needs to use the page. That brings us from notes and memory to the buttons and forms people already rely on.

Before we get to websites, there's another sort of note that has to be right: the one showing what a system costs. In my system, an AI model had no price listed. Every call showed a cost of zero, even as the real bill kept growing.

A spending limit watched that meter, but zero never crossed the limit. So the safeguard protected nothing. It was like watching a fuel gauge stuck on full while the tank emptied.

Now every run checks the account balance first. It refuses a run the balance can't cover, reports the real spend afterward, and treats a zero reading as a bug until proven otherwise. That rule is plain, but it matters: a check only helps when the number it checks reflects what's happening.

In late August, OpenAI said ChatGPT's agents can now work inside websites you're signed in to. You can hand off a task and walk away while the agent keeps going. That makes the way a site is built part of whether the task gets done.

An agent uses a page through its structure: the labels, buttons, and form fields it can find by name. A form field with no label can leave it guessing. A button that's only a picture may give it nothing clear to press. If a step depends on reading an image, it can get stuck there too.

You can picture it like giving someone directions inside a building. A clear sign points to the right door. A door with no sign leaves them trying handles, and the agent has less room to work around that kind of confusion.

When an agent gets stuck on your site, the task goes to a site where it doesn't. That's why clear pages matter in both senses: buyers need answers they can use, and software needs parts it can identify.

Across all four essays, the same practical idea keeps coming back. A lesson should reach the next session, a meter should report the real spend, and a page should make its purpose clear to the person or agent using it.

You can read the essays on bobmichaels.ai. I'm Bob Michaels. Thanks for hanging out. See you in two weeks.