> ## Documentation Index
> Fetch the complete documentation index at: https://docs.agentbees.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Browser agent

> Give a task a real, sandboxed web browser — navigate, fill forms, screenshot, and verify pages — and watch it work live.

Some work needs the **web**, not just the repo: checking that a deployed page
renders, logging into an app to reproduce a bug, pulling facts off a site, or
filling out a form. The **browser agent** gives a task a real web browser it can
drive — and streams what it's doing back to you **live**.

## What you can do with it

Turn it on for a task and the agent can browse the web as part of its work:

* **Verify a deployed site** — "open the staging URL and tell me whether the
  pricing page renders and the CTA works."
* **Reproduce a bug behind a login** — navigate an authenticated app and capture
  exactly what the user sees.
* **Read a page** — "open this article and summarize the top items," or grab a
  page's heading, prices, or table.
* **Drive a flow** — fill a form, click through a multi-step wizard, follow links.
* **Research across sites** — gather and compare information from several pages in
  one task.

## Watch it live — and keep the screenshots

<Frame caption="The live Browser panel — the agent's browser streams in real time next to the chat/terminal, with the screenshots it captured below.">
  <img src="https://mintcdn.com/agent-bees/RyHDU_9MX4TN4NTC/images/browser-agent.png?fit=max&auto=format&n=RyHDU_9MX4TN4NTC&q=85&s=56011e23616063c9e67710d410696d6c" alt="The live Browser panel in AgentBees showing the agent's browser and captured screenshots" width="2216" height="1432" data-path="images/browser-agent.png" />
</Frame>

* A **Browser panel** shows the agent's browser **as it happens** — every page it
  opens, live, whether you're in the chat view or the developer view.
* Each navigation is **captured as a screenshot** you can click to preview.
* The screenshots are **saved with the task**, so they're still there when you
  come back later — reopen the panel any time from the **🌐 Browser** chip.

## Two backends — pick isolation or speed, per task

Choose the browser backend for each task from the **New Task** form:

| Backend     | Where the browsing runs                        | Best for                                   |
| ----------- | ---------------------------------------------- | ------------------------------------------ |
| **🌐 Grid** | Your own infrastructure — an isolated Chromium | Private, authenticated, or sensitive flows |
| **rtrvr**   | rtrvr's managed cloud (fast)                   | Public, non-sensitive pages                |

<Warning>
  With **Grid**, browsing stays inside your own network — nothing leaves. **rtrvr** is faster, but
  the page content is sent to rtrvr's cloud, so use it only for **public, non-sensitive** flows.
</Warning>

## Turn it on

<Steps>
  <Step title="An admin enables it for the org">
    In **Admin → Agent LLM**, turn on the browser agent. For the rtrvr backend, add the org's rtrvr
    API key in the same panel (Grid needs no key).
  </Step>

  <Step title="Pick a backend on the New Task form">
    Under **Advanced options**, set **🌐 Browser agent** to **Grid** or **rtrvr**.
  </Step>

  <Step title="The agent gains web-browsing tools">
    It navigates, screenshots, reads, and fills pages on its own — and you watch it in the Browser
    panel.
  </Step>
</Steps>

<Note>
  The browser agent is **off by default** — browsing is an explicit opt-in, both at the org level
  and per task, so a task only reaches the web when you choose it.
</Note>