
Browser agent: AI that browses and acts on websites
In short: A browser agent is an AI agent that navigates websites, reads pages and fills forms for you, which exposes it to instructions hidden in web content.
A browser agent is an AI agent that controls a web browser to complete tasks: it opens pages, reads their content, clicks links, fills forms and moves between sites. It usually works from the page's structure (the DOM or accessibility tree), screenshots or both. It also inherits every risk of reading untrusted web content, which OWASP's 2025 list calls indirect prompt injection.
Browser agent in plain words
A very fast assistant with your browser open, who reads every page literally, including the parts written to trick it.
How it works
The agent receives a simplified view of the page: visible text, interactive elements with IDs, sometimes a screenshot. It chooses an action such as "click element 14" or "type into the search box", the browser driver executes it, and the new page state comes back. Automation libraries such as Playwright often sit underneath.
The WebMCP proposal describes today's approach as scraping the DOM, taking screenshots and simulating clicks, and offers an alternative: pages register JavaScript tools that an agent can call directly. Its specification, read in November 2026, is a Draft Community Group Report from the Web Machine Learning Community Group and is not a W3C standard.
Example
As an illustration, "find three flights to Delhi under a set budget next Friday and list them": the agent opens a travel site, sets the date and city, sorts by price, reads the results and returns a table. It stops before booking because payment needs a person.
Where you'll meet it
- Browser extensions that let an assistant act in your own browser
- AI browsers with built-in agents
- QA and end-to-end test generation
- Sites experimenting with WebMCP tools for agents
What it changes for you
For web teams, agents are a new kind of visitor: clear labels, stable forms and, eventually, explicit agent-facing tools make your site usable by them. For security teams, any page the agent reads can carry injected instructions. OWASP's mitigations are a practical starting point: limit the agent's access to what the task needs and require human approval for high-risk actions.
Browser agent vs web scraper
A scraper extracts data with fixed rules you wrote. A browser agent decides its own steps and can act as well as read.
Common mistake
My view: letting a browser agent loose in a profile signed into email, bank and work tools at once. That is the lethal trifecta in one window. Use a separate profile with only what the task needs.
Related terms
Computer use · WebMCP · Prompt injection · Lethal trifecta · AI agent
From the archive: A GET-only browser still needs a write boundary · A browser can use less CPU and still make the agent wait longer
Give a browser agent its own profile with only what the task needs.


