Agent Experiences

Muse and Dots Both Drive a Browser

By Alan WeibelPublished 8 min read

Meta Muse

Meta launched Muse on September 8, 2026. It is a personal AI agent that can, in Meta's own words, "open a browser, fill out forms, and negotiate on their behalf." Each person's Muse runs in its own virtual machine (a separate computer simulated in software), and a second agent called Sentinel decides what is allowed out to the internet. On September 29 Meta opened it to small businesses, with connections to Shopify, Slack, Asana, Intuit, and Canva.

Source: Meta Newsroom, September 8, 2026

OpenAI Dots

OpenAI launched Dots at its DevDay event on September 29, 2026, and described them as "remarkably capable, always-on agents built to handle everything." Each Dot runs on GPT-6 Astra, gets its own computer and browser in the cloud, and can reach more than 4,000 apps through ChatGPT plugins. Dots are available to ChatGPT Pro and Business Premium subscribers in eligible markets.

Source: TechCrunch, September 29, 2026

Much of what we have written about websites comes back to one question: can an AI tool read your page and quote it correctly? Muse and Dots raise a different question, because both of them open a browser and use your site much as a person would.

That is a bigger change than it sounds. A crawler (a program that fetches pages so a search engine or AI tool can use them later) grabs your page, pulls out what it can, and leaves. An agent with a browser lands on your page, looks for the thing it was sent to do, and tries to do it. It fills in the form, gets through checkout, cancels the subscription, or checks your price against four competitors it has open in other tabs.

What Meta and OpenAI actually launched

Meta's description is the blunter of the two. Muse can "open a browser, fill out forms, and negotiate on their behalf," and it "keeps working after people close the app." Meta built it so each person gets a dedicated virtual machine, with Sentinel deciding what is allowed out to the internet.

OpenAI's Dots work in a similar way. Each Dot gets its own cloud computer and its own browser, connects to more than 4,000 apps, and keeps running in the background instead of only during a chat. The user decides what a Dot may do on its own, what it has to ask about first, and what it must never do. Changing a password always stays with the human.

Neither company has published usage numbers, so nobody outside Meta and OpenAI knows how many people actually run these agents. Take the launches as a sign of where things are heading, not as proof that lots of people already use them.

Why this lands differently than crawling did

While AI tools were mostly crawling pages and citing them, the worst case was fairly gentle: you got left out. If a crawler couldn't read your page, you didn't appear in an answer. That hurt, but it happened quietly.

An agent that uses your site fails much more visibly. It arrives carrying someone's intention and their payment details. If your checkout breaks while the agent is using it, your customer doesn't notice a missing citation. Your customer hears that their assistant couldn't finish the job on your site, while the same job worked on a competitor's.

The legal ground has shifted under all of this too. On August 4, 2026, the Ninth Circuit, a US federal appeals court, set aside the order that had stopped Perplexity's Comet browser from shopping on Amazon. The court reasoned that an agent following a user's instruction is the user acting. So blocking agents that act for a user now looks, legally and practically, a lot closer to blocking the customer.

What changes for the people using them

Three things change, and the third deserves the most thought.

Sessions get longer and less supervised. Someone asks for a result, closes the app, and comes back later to a finished task. Your analytics will probably record a visit that looks nothing like a human one. The agent won't stop to read your headline, and a pop-up meant to persuade won't persuade it.

Comparing gets cheap. An agent can hold your page and four competitors' pages open at the same time and weigh them against each other. If your page is even a little harder to make sense of than a rival's, that now costs you.

And your customer is no longer in the room. Everything you built to persuade, from the testimonial to the urgency banner to the carefully ordered landing page, is aimed at someone who isn't looking. What reaches them is whatever the agent reports back.

What to do now

Start with the plain, unexciting work, because an agent can't act on a page it can't read. Content that only appears after JavaScript runs in the browser, a login wall in front of something that should be public, a form that needs you to drag something into place: each of these used to cost you a citation, and now it costs someone a finished task. Our what gets cited page sorts out which of these problems have evidence behind them and which are folklore.

Then look at the parts of your site where something gets done: checkout, booking, account changes, and support requests. Try each one using only the keyboard, without relying on the layout, which is closer to how an agent gets around.

Decide on purpose which automated visitors you let in, instead of leaving it to a default setting. Search crawlers (the ones that build a search index), training crawlers (the ones that collect text to train AI models), and agents acting for a user are three separate decisions. A single blanket rule in robots.txt (the file that tells automated visitors which parts of your site they may fetch) almost always gets one of them wrong. Our AI opt-out controls guide has a template for each choice.

If you want to go further than making your pages easy to read, WebMCP (a proposed way for a web page to offer agents named actions they can call directly) is the first serious attempt at letting a page hand an agent ready-made tools, instead of making it pretend to be a person with a mouse. It entered a Chrome origin trial (an early test that websites can opt into) in June 2026. It hasn't shipped, and the specification is still changing. Learn how it works now, but wait until it ships before building anything you depend on.

What nobody can tell you yet

Three things are still unknown: whether either product will last, whether these agents will name themselves honestly in their user agent (the name a browser or bot sends with every request to say what it is) or arrive looking like an ordinary copy of Chrome, and whether the economics work for anyone except the companies that build the AI models.

Whatever happens, check whether an agent can finish a task on your site.

See the path on your own site

Muse and Dots work your pages through a browser. See the route an agent takes through one of yours, and the step where it gives up.