LangChain adds a hosted browser and sign-in handoff to Managed Deep Agents

The default session setting gives each conversation a fresh browser window, without cookies or logins from another conversation. An optional user-scoped setting shares cookies and other browser state across a signed-in user’s conversations, allowing a later conversation to reuse a sign-in.

Why LangChain chose Stagehand

LangChain built the tools on Browserbase’s open-source Stagehand SDK. The browser_snapshot tool reads a page’s accessibility tree, a text view that labels interactive elements; browser_run acts on the page through batches of actions or Playwright-style code; and browser_screenshot captures the rendered page. A private link between the browser and agent sandboxes lets the browser reach an app running in the agent’s sandbox for testing, Runkle wrote.

In a Nov. 3 follow-up, Runkle explained that LangChain compared seven browser toolsets, examined LangSmith traces and evaluated changes to the tools. In LangChain’s evaluation of 100 WebVoyager tasks using Claude Sonnet 5.5, the Stagehand tools passed 90% for $1.98 total, with a median of 17.4 seconds per task. Playwright MCP passed 78% for $26.71 in that comparison. Runkle reported that Stagehand’s run tool averaged 10 steps per task, compared with 26 model turns for Playwright MCP. LangChain’s results compare those toolsets on the tasks it tested, not browser performance across every website.