LangChain's Deep Agents can now click through sites with no API
Managed Deep Agents v0.10 adds browser tools so an agent can fill in forms and click buttons on sites with no API, plus subagents that keep working in the background instead of blocking the conversation.
LangChain shipped Managed Deep Agents v0.10 in public beta on 29 October. The headline feature is browser use: four new tools, browser_snapshot, browser_run, browser_screenshot and browser_handoff, that run in LangChain’s own sandbox so an agent can read a page’s accessibility tree, click and fill fields, take screenshots, and hand the open window to a person when it hits a login screen. By default each conversation gets its own browser with no cookies left over from other threads, though setting scope="user" shares one signed-in browser across a person’s threads instead.
The release also lets subagents run asynchronously. Adding "async": True to a subagent’s definition starts it on its own thread, so the orchestrator keeps talking to the user instead of blocking on the result:
"subagents": [{
"name": "researcher",
"async": True,
}]The release also removes the old connectors.mcp() API: MCP servers now have to be declared in tools/mcp.ts or tools/mcp.py.
Why it matters: a lot of real tasks, cancelling a subscription, filing a claim, booking a flight, only exist behind a login form with no API, and curl just returns markup a browser has not run yet. Built-in browser tools save every team from bolting their own Playwright script onto an agent, and non-blocking subagents mean a background research task no longer freezes the conversation while it runs.
The caveat: Managed Deep Agents is a separate hosted product from the open-source deepagents package, and the breaking changes here will need fixing in any agent before it can upgrade.