Muse and Dots Both Drive a Browser

Since this site started, the question it has been answering is whether an AI system can read your page and quote it correctly. Muse and Dots ask a different one. Both of them open a browser.

That is a change in kind, not degree. A crawler fetches your HTML, extracts what it can, and leaves. An agent with a browser lands on your page, finds the thing it was sent to do, and tries to do it: fill the form, complete the checkout, cancel the subscription, compare the price against four competitors it has open in other tabs. Reading is now the easy half.

What actually shipped

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 to run in a dedicated virtual machine per person, with a second agent called Sentinel deciding what is allowed out to the internet.

OpenAI's Dots are structured similarly and aimed slightly differently. Each Dot gets its own cloud computer and its own browser, connects to more than 4,000 apps, and runs continuously rather than per-conversation. The user sets what it may do alone, what it must ask about first, and what it must never do. Changing a password stays with the human.

Neither company has published usage numbers, so nobody outside them knows how many people are actually running these. Treat the launches as a signal about direction, not as evidence of scale.

Why this lands differently than crawling did

The crawl-and-cite era had a comfortable property: the worst case was being left out. If a crawler could not read your page, you did not appear in an answer. Bad, but quiet.

An agent that operates your site has a louder failure mode. It arrives holding a person's intention and their payment method. If your checkout breaks under it, the person does not see a missing citation. They see a task their assistant could not finish, on your site, while a competitor's worked.

The legal ground shifted underneath this too. On August 4, 2026 the Ninth Circuit vacated the order that had blocked Perplexity's Comet browser from shopping on Amazon, reasoning that an agent acting on a user's instruction is the user acting. Blocking user-directed agents is now much closer, legally and practically, to blocking the customer.

What changes for the people using them

Three things, and the third is the one worth sitting with.

The session gets longer and less supervised. Someone asks for a result, closes the app, and comes back to a finished task. Your analytics will show a visit that looks nothing like a human's, because it is not pacing itself, not reading your hero copy, and not susceptible to anything you put in a modal.

Comparison gets cheap. An agent can hold your page and four competitors open simultaneously and reconcile them. Being slightly harder to parse than a rival is no longer a small disadvantage.

And the person is no longer in the room. This is the part most sites have not priced in. Everything you built to persuade — the testimonial, the urgency banner, the carefully sequenced landing page — is aimed at someone who is not looking. What reaches them is whatever the agent reports back.

What is worth doing now

Start with the unglamorous half, because an agent cannot act on a page it cannot read. Content that only appears after JavaScript runs, a login wall in front of something that should be public, a form that depends on a drag gesture: each one is now a task failure rather than a missed citation. Our what gets cited page covers which of these are evidenced and which are folklore.

Then look at the parts of your site that complete something. Checkout, booking, account changes, support requests. Try them without a mouse, without JavaScript, and without reading the visual layout — that is roughly the experience an agent has. Most sites discover at least one step that only works because a human guessed what a designer meant.

Decide your crawler posture deliberately rather than by default. Search crawlers, training crawlers, and user-directed agents are three separate decisions, and a single blanket rule in robots.txt almost always gets one of them wrong. Our AI opt-out controls guide has the templates for each posture.

If you want to go further than being readable, WebMCP is the first serious attempt at letting a page offer an agent named tools instead of making it simulate a person with a mouse. It entered a Chrome origin trial in June 2026, it is not shipped, and the specification is still changing. Worth understanding, too early to build a business on.

What nobody can tell you yet

Whether either product lasts. Whether these agents will identify themselves honestly in their user agent, or arrive looking like Chrome. Whether the economics work for anyone but the model providers.

What is already true, regardless of how those resolve, is that the bar moved. A page that an agent can read is table stakes. A page an agent can finish a task on is the thing worth checking, and most sites have never been tested for it.