What are browser agents? AI that clicks, types and browses for you

What a browser agent does on each step

Give a browser agent the job "find the number of stars of the browser-use repo". It starts a browser, opens a search or the repository page, reads what is there and decides the next click. A language model picks each action. A small program carries it out in a real browser. The result goes back to the model so it can pick the next one.

The model never touches the browser itself. OpenAI's computer use guide describes three ways to wire this up. The model gets screenshots and answers with mouse and keyboard actions. Or it gets page text and controls as function tools. Or the vendor hosts the browser. The wiring a product uses decides what the agent can see and where it can go wrong.

If the word agent is new, the AI agent lesson explains the loop in general. A browser agent is that loop with a browser as one of its tools.

Five ways to get a browser agent

Each row names a product or project and where its browser runs, as the vendor's own pages describe it on 6 November 2026. The first two kinds take a task in plain language, while the last three need code or a settings file.

Lesson 2 walks through the first row with ChatGPT agent, and lesson 3 covers the last row. Playwright MCP is an MCP server, a plug that gives an AI client a tool. The MCP lesson shows how a client connects to one.

Run a first browser agent in five steps

The quickest way to see the loop is to run it. Browser Use is an open-source Python library, and its README gives a short quickstart. You need Python 3.11 or newer, the uv package manager and an OpenAI API key. The BROWSER_USE_API_KEY line in the sample is optional. OpenAI bills the model calls, so keep the first task short.

import asyncio

from browser_use import Agent, Browser, ChatBrowserUse, ChatOpenAI
from dotenv import load_dotenv

load_dotenv()

async def main():
    llm = ChatOpenAI(model='gpt-5.6-luna', reasoning_effort='xhigh')
    # llm = ChatBrowserUse(model='bu-2-0')  # Use BU2 instead; requires BROWSER_USE_API_KEY
    agent = Agent(
        task="Find the number of stars of the browser-use repo",
        llm=llm,
        # browser=Browser(use_cloud=True),  # Use a cloud browser; requires BROWSER_USE_API_KEY
    )
    history = await agent.run()
    print(history.final_result())

if __name__ == "__main__":
    asyncio.run(main())

1. Make a new folder and run uv init --python 3.12 in it, then uv add browser-use.
2. Create a file named .env and put your key in it as OPENAI_API_KEY=your-key.
3. Save the code below as agent.py. The task is one sentence that names an end state you can check.
4. Run uv run agent.py. The README says the agent opens a browser, looks up the repository and prints its answer.
5. Open the repository page yourself and compare the number with what the agent printed.

Code from the browser-use README, read on 6 November 2026. The two commented lines swap in Browser Use's own model or a cloud browser, and both need a Browser Use API key.

How an agent sees the page: a picture or the structure

The three wirings split into two ways of seeing a page. A screenshot agent gets a picture and answers with coordinates, the way you'd point at a screen. A structure agent gets the page as text, with each button and link listed. Playwright MCP is built on the second. Its README says it works through accessibility snapshots, needs no screenshots or vision model, and avoids the ambiguity of screenshot-based approaches.

Anthropic's browser use demo argues the same against pixel coordinates, which break when a window changes size. Screenshots are the way in when the target is a desktop program. OpenAI's sample app drives a desktop with PyAutoGUI, and Anthropic's computer use demo drives a Linux desktop in Docker.

Names changed fast, so date what you read

OpenAI launched ChatGPT agent in July 2025 and its Atlas browser in October 2025. Wikipedia lists Atlas as shut down on 9 August 2026 and ChatGPT Work as launched in July 2026. A tutorial written earlier in the year can name a menu that is gone. The timeline lists the dated changes this course relies on.

Common mistakes

• Trusting the agent's last message. OpenAI's sample app README says a final answer does not prove the task succeeded, so check the page.
• Starting with a job that spends money or sends a message. Anthropic told developers to begin with low-risk tasks when it opened computer use in 2024, and a read-only job on a public page is the lowest risk.
• Handing over your everyday signed-in browser for a job that never needed an account.
• Expecting a CAPTCHA, the puzzle a site uses to check that you are a person, to be solved. The browser-use README says no browser configuration guarantees that every CAPTCHA can be avoided or solved.
• Following a tutorial's menu names without checking its date against the vendor's current docs.

Try it yourself

• Write one job you do in a browser every week as a start page plus an end state, such as open one shop's delivery page and report the free delivery threshold.
• Find that job in the table: which kind would you try first, and which of your accounts would it need?
• Run the browser-use quickstart with a task of your own on a public page, then compare its answer with the page.
• Find the date on a browser agent tutorial you used before and check whether its menu names still appear in the vendor's docs.

Next: hand a browser task to ChatGPT

In the ChatGPT agent tutorial, lesson 2 gives ChatGPT agent a browser task, from writing it to the approval prompts, using OpenAI's system card.

Frequently asked questions

Sources

1. OpenAI: Computer use guide (read November 6, 2026)https://developers.openai.com/api/docs/guides/tools-computer-use
2. OpenAI: ChatGPT agent system card (read November 6, 2026)https://deploymentsafety.openai.com/chatgpt-agent
3. openai/openai-cua-sample-app README (read November 6, 2026)https://github.com/openai/openai-cua-sample-app
4. browser-use/browser-use README (read November 6, 2026)https://github.com/browser-use/browser-use
5. microsoft/playwright-mcp README (read November 6, 2026)https://github.com/microsoft/playwright-mcp
6. anthropics/anthropic-quickstarts: computer use demo README (read November 6, 2026)https://github.com/anthropics/anthropic-quickstarts/tree/main/computer-use-demo
7. anthropics/anthropic-quickstarts: browser use demo README (read November 6, 2026)https://github.com/anthropics/anthropic-quickstarts/tree/main/browser-use-demo
8. Anthropic: Computer use, Claude 3.5 Sonnet and Claude 3.5 Haiku (read November 6, 2026)https://www.anthropic.com/news/3-5-models-and-computer-use
9. Anthropic: Mitigating prompt injections in browser use (read November 6, 2026)https://www.anthropic.com/research/prompt-injection-defenses
10. Wikipedia: ChatGPT Atlas (read November 6, 2026)https://en.wikipedia.org/wiki/ChatGPT_Atlas
11. Wikipedia: ChatGPT (read November 6, 2026)https://en.wikipedia.org/wiki/ChatGPT
12. Wikipedia: Claude (AI) (read November 6, 2026)https://en.wikipedia.org/wiki/Claude_(AI)