LangChain adds managed browser to Deep Agents for web tasks

Why agents need a browser

LangChain argues that much of the work people want agents to do happens in a browser. Booking travel, canceling a subscription or filing a claim take place on sites built for people clicking through pages. Often there is no API an agent can call, and fetching a page with curl returns the HTML before any JavaScript renders it and cannot click anything.

The company points to personal agents such as Instinct and Meta's Muse, which already work this way. It also cites DataDome, which saw agentic browser requests jump 4.7x in September alone.

LangChain names two common use cases so far. The first is actions an API can't cover, such as booking a flight or a restaurant reservation. When a site does have an API for the task, the agent should use it. The second is end-to-end testing, for example checking a frontend change before it merges. LangChain says it uses this in Open SWE, its coding agent.

Setup, sessions and login handoff

Developers add a browser by exporting one from browser.py at the project root, using the define_browser function. The scope setting controls how sessions are shared:

• thread (the default): each thread gets its own browser, with no cookies or logins left over from another thread.
• user: one browser is shared across a signed-in user's threads. Each thread has its own window, but cookies and other state are shared.

The browser stops after idle_ttl_seconds without activity, 10 minutes by default, so users do not pay for an unused browser.

For steps that need a person, such as signing in, the agent calls browser_handoff with a reason. In LangSmith Studio, the user gets an interactive card with the same browser, signs in personally and clicks Hand back. The agent then continues already signed in. This keeps the password out of the model's messages, traces and the model provider. Handoff comes with every managed browser and works automatically in channels such as Slack. With user scope, the sign-in carries over to later threads.

How the browser and tools work

Managed Deep Agents runs the browser in its own LangSmith sandbox, separate from the agent, so nothing runs on the user's machine. LangChain says it also has a private link to the agent's sandbox, which makes end-to-end testing possible.

The agent gets four tools:

• browser_snapshot reads the page's accessibility tree and labels interactive elements with IDs.
• browser_run acts on the page, with a batch of actions like click and fill or with JavaScript using a Playwright-style page object, which Stagehand translates into its own commands.
• browser_screenshot captures the rendered page.
• browser_handoff passes the browser to a person.

Because browser_run takes code, the agent can write a whole chain of steps as one call and return only what it needs. LangChain says this keeps long tasks fast and most of the page out of the agent's context. The tools are built on Stagehand, Browserbase's open source SDK.

On 100 tasks from WebVoyager, a benchmark of everyday tasks on live websites, the tools passed 90% for $1.98 in total, according to LangChain. That was the highest pass rate of the seven toolsets it tested. The closest, Browserbase's MCP server, passed 89% at 9x the cost.

End-to-end testing and what it means

For testing, the app runs in the agent's own sandbox and the browser reaches it through a localhost bridge. When the browser opens localhost:5173, the request travels over a private link to port 5173 in the agent sandbox, so the app needs no public URL. The agent starts the app, clicks through flows like a signup form and checks what renders. After a passing test, it can save its steps as a test file, since each browser_run call is a piece of an ordinary Playwright script. Later runs replay that file with no model calls until the page changes.

For developers, LangChain says giving a managed agent a browser takes a few lines of code. Agents built with open source Deep Agents can also use a browser, with users bringing their own, such as a hosted one from Browserbase. LangChain plans a deep dive next week on how it tested and tuned the tools.