Skip to content

what is an agentic browser

github-actions[bot] edited this page Nov 13, 2026 · 4 revisions

What is an agentic browser?

An agentic browser is a browser that can carry out a task for you, rather than only display pages while you do the clicking. You give it an instruction in words, and it navigates, reads, clicks, fills forms and reports back. The browser is still a browser: what is new is that an AI agent sits inside it with the ability to drive the same UI you would have driven.

That is the whole definition. Everything else on this page is the part the definition hides, and the reason people search this term twice.

The term covers two different products, and that is why it is confusing

Search this phrase and you get two kinds of result that barely overlap: consumer browser reviews, and security advisories. They are not talking past each other by accident. There really are two things.

A browser you use, with an agent inside it. You install it, you log in, you browse. When you want something done, you ask. Perplexity's Comet, OpenAI's ChatGPT Atlas, Opera Neon, Fellou, Dia, Sigma and the open-source BrowserOS all sit here, as do the agent features being folded into Chrome and Firefox. The agent inherits your session, your cookies and your logged-in accounts, because it is running in your browser.

A browser your code drives, with an agent in the loop. No window you sit in front of. A program starts a browser, an LLM decides what to do next, and the browser does it. browser-use, Skyvern, Stagehand and invisible_playwright_mcp - this wiki's project - are this shape. Nobody is logged in unless you arrange it.

The words are almost the same and the risk profile is opposite. If you are comparing tools, work out which one you meant first, because a review of Comet tells you nothing useful about a library, and vice versa. This wiki keeps that split on its own page: AI browser vs AI browser agent.

The four things any agentic browser has to solve

Whichever shape it is, the same four problems show up, and how a product answers them is a better comparison axis than its feature list.

How the agent perceives the page. Two families. Either the model reads a structured description of the page - the accessibility tree, or the DOM reduced to something a model can hold - or it looks at a screenshot and works in pixels. Structured reading is cheaper and more reliable when the page is well built; pixels are the fallback for canvas, custom widgets and anything the tree does not describe. Most serious tools now do both and pick per step.

How it acts. A selector-based click is precise and brittle. A coordinate click is general and blind. Whether the resulting event looks like a person depends on the layer underneath, which is a separate question from the agent.

What it is allowed to do. This is the one that will decide the category. An agent that inherits your logged-in session can do anything you can do, including things you did not ask for, if it is talked into it by text on a page.

Whether the site can tell. An agentic browser is still automation as far as a defended site is concerned, and the tells live in the browser build and the network path, not in the agent's reasoning. This wiki has a whole section on that: when the agent gets blocked.

The security argument, stated fairly

The reason a security vendor outranks a product page for this term is that the consumer shape has a genuinely unsolved problem: prompt injection. The agent reads the page in order to act on it, so text on the page arrives in the same channel as your instruction. A page can contain a sentence addressed to the agent rather than to you. If the agent is browsing with your session, the worst case is not a wrong answer, it is an action taken with your credentials.

The research literature in 2026 is actively proposing browser-level answers to this - work on a same-origin policy for agentic browsers is one direction, the idea being that an agent's authority should be scoped the way a script's already is. None of it is settled. Treat any product claim of "we handle prompt injection" as a claim, and ask what it means mechanically.

The library shape dodges most of this by default, for a boring reason: it usually starts with no session at all. That is not virtue, it is scope. The moment you hand a code-driven agent a logged-in profile, you have the same problem, which is why this wiki has a page asking whether you should log your agent into accounts before you do it.

Do you want one?

  • You want to save yourself clicks on sites you are logged into. A consumer agentic browser. Accept that it acts with your session, and read what it does with your data before you enable it on your bank.
  • You want a task to run repeatedly, unattended, on a schedule. Not a consumer browser. You want the library shape, where the run is a program you can log, replay and rate-limit.
  • You want to pull data out of sites at volume. Probably not an agent at all, at least not for the whole job. See AI browser agents vs traditional scraping: an agent is the right tool for the part that changes and the wrong tool for the part that repeats.
  • You want your existing assistant to be able to open a page. That is the MCP route, and it is the cheapest of the four to try: the MCP server.

Short answers to the questions that lead here

What does agentic mean in a browser? That the browser can take a sequence of actions toward a goal you stated once, instead of executing the single action you just clicked.

Is an agentic browser the same as an AI browser? Not quite. An AI browser may only summarise and answer; an agentic one acts. Products blur the line because acting is the feature people pay for.

Is an agentic browser safe? The mechanism to worry about is prompt injection, not the AI generally. The exposure scales with what the agent is logged into, so the honest answer depends entirely on how you configure it.

Are there open-source agentic browsers? Yes, in both shapes. Open-source agentic browsers and agents covers what exists and what each one actually gives you.

Will sites block an agentic browser? Defended sites block automation whether or not there is an agent involved, and they judge the browser and the network path, not the intent. Start with why an agent gets blocked.

See also: Choosing an AI browser agent for the decision framework, What is an AI web agent? for the library shape in detail, and the MCP server for turning an assistant you already run into one.

Sources


Written while maintaining invisible_playwright_mcp, which is one of the tools in the second category. The page is deliberately about the category rather than the product, and the section that tells you when you do not want an agent is there because it is the honest answer more often than a project in this space likes to admit.

Alternatives and Comparisons

When the Agent Gets Blocked

Using the Agent

Clone this wiki locally