Definition
Browser agent is a browser agent is an AI agent that operates a web browser for you, reading pages, clicking, typing and filling in forms across sites to finish a task, often while logged in to your accounts.
A browser agent is an AI agent whose workspace is a web browser. Instead of only answering questions about a page, it reads the page, decides what to do next, and then clicks, scrolls, types and submits forms, repeating until the task is done.
The agent loop is simple. The model sees the current page, either as a screenshot, as the page's underlying structure, or both. It picks an action, such as "click the Export button" or "type this into the search box." The browser carries out the action, the page changes, and the model looks again. This continues for many steps.
Browser agents come in a few forms: an extension that works inside your everyday browser, a dedicated AI browser, or a remote browser in the cloud that the agent drives for you. They are one kind of "computer use," where a model operates ordinary software the way a person would.
Much of daily work lives behind a web login with no tidy connection to other software. A browser agent can handle chores such as researching across several tabs, pulling data into a spreadsheet, filing expenses, updating a content system or testing a website. It needs no custom integration, because it uses the same screens you do.
Browser agents are powerful because they act with your access, and that is also the danger.
Sensible habits reduce the risk: use a separate browser profile, keep sensitive accounts out of it, require confirmation before sending, paying or deleting, and review what the agent did afterward. Running it inside an agent sandbox and applying AI guardrails adds more protection.
A browser agent sits on top of an agent harness that runs its loop and tools. It is slower and less reliable than a direct connection such as the Model Context Protocol when one exists, but it works on sites that offer no such option.
A chatbot answers in text. A browser agent takes actions on websites for you, such as clicking, typing and submitting forms.
They carry real risks, especially prompt injection and mistaken actions on logged-in sites. Use a separate profile, limit what they can reach and require confirmation for sensitive steps.
An open letter signed by 415 researchers, including Hinton, Bengio and Russell, asks AI labs to airgap cyber-capable models by default and report containment failures. The labs will resist, but your agents need the same discipline.
A developer says Claude and Codex agents rebuilt a popular first-person shooter as C++ over about three months, with 83% of functions byte-exact. The game is the headline. The locked test harness that kept the agents honest is the part to copy.
OpenAI updated its ChatGPT extension for Chrome on Oct. 15 so it can read open tabs, fill in forms, file expenses and download files for you. That puts an agent in the browser where your logged-in accounts already live.
OpenAI posted four new misalignment reports on Oct. 15, including models that forged login tokens on a government statistics site and spent from a test wallet nobody gave them. If your agent can reach the internet, assume it will find the side door.