OpenAI's ChatGPT extension can click buttons. Here's why that's not the same problem Manifest solves.

OpenAI shut down Atlas, its standalone browser, on August 9, 2026. What replaced it is a Chrome extension that reads the page you're on, plus "Agent Mode," which can click buttons and fill out forms on its own. A few people have asked me if this makes Manifest redundant. It doesn't, and the reason why is a good excuse to explain what Manifest actually does that a browser agent doesn't.

What OpenAI shipped

Two separate things, worth keeping apart:

The extension is a sidebar. It sees the page you're looking at, so you can ask it questions, get a summary, or kick off a longer task without copying text into a separate tab. It's the reading layer.

Agent Mode is where the actual clicking happens, and the mechanism matters: it works by taking a screenshot of a virtual browser window, looking at that image to figure out where the buttons and fields are, clicking, then taking another screenshot to see what changed. It's not parsing the page's structure — it's looking at a picture of the page, the same way you would, and deciding what to do next.

That's a genuinely impressive piece of engineering. It's also a completely different approach to "how does an agent know what it can do on this page" than the one Manifest takes.

The problem with figuring it out from a screenshot

Screenshot-and-guess works, until the page changes. A button moves six pixels to the left after a redesign, a modal renders half a second later than expected, a form field is disabled until three other fields are filled — and now the agent is re-deriving the whole layout from scratch, on every single step, from an image. It's slow (a vision pass per action), and it's fragile in a specific way: there's no persistent, structured record of what's required before what. The agent's understanding of "this field needs that field filled first" lives inside a language model's guess about a screenshot, not in a graph it can check.

I don't say this to knock Agent Mode — for a general-purpose assistant taking over your browser for one task at a time, screenshot grounding is a reasonable default. It's built to work on literally any site, with no integration required, for a person who just wants something done. That's a different job than the one Manifest is built for.

What Manifest does instead

Manifest doesn't take screenshots and it doesn't guess. It visits a URL, reads the page's accessibility tree the same way a screen reader would, cross-references it against the actual DOM for the details accessibility trees miss — required fields, input types, disabled states — and returns a structured JSON manifest: every button, form, and input on the page, each with a resolved locator and a requires field that says what has to happen before that action is valid.

The difference isn't "better at clicking." It's that the sequencing logic — what's gated on what — is data your agent can read once and reason over, instead of something re-inferred from pixels every time the page reloads. An agent built on Manifest doesn't need a vision pass to know the submit button is blocked until the email field is filled; it's already in the requires array.

Where this actually matters

If you're a person asking ChatGPT to book a flight, Agent Mode is exactly the right tool — you don't want to install anything or write code, you want it to just go do the thing while you watch.

If you're building your own agent — a LangChain tool, an MCP-connected loop, a custom Playwright script that needs to reason about a page — you're not going to route that through someone else's chat product. You need something you can call from code, that returns a data structure you can reason over, on whatever model you're already using. That's the actual gap Manifest sits in, and it's a gap regardless of how good OpenAI's screenshot-based clicking gets, because it's not solving the same problem: Manifest is infrastructure other agents call into; Agent Mode is a closed product OpenAI built for its own chat interface.

What this doesn't mean

There's no pretending there's zero overlap here. If OpenAI (or anyone else) ever exposes their page-understanding step as a public API — not the whole agent, just "give me the structured actions on this page" — that would compete with Manifest directly. Nobody's done that yet, as far as the public record shows. What's shipped so far is a consumer feature bundled inside one company's chat app, not a primitive other developers can build on.

More agents doing real work on the web, from any lab, is a reason for more infrastructure like this to exist — not less.