LangChain Managed Deep Agents v0.10 adds browser use

Most of the software a small business runs has no API. The vendor portal, the insurance claim form, the supplier's order page. On October 29, LangChain shipped Managed Deep Agents v0.10 with browser use, so agents can work those sites directly. The part we like more: a login handoff that keeps your passwords out of the model entirely.

What actually happened

Per LangChain's release post, v0.10 is in public beta on LangSmith. You enable the browser with one root file:

from managed_deepagents import define_browser

browser = define_browser(scope="thread", idle_ttl_seconds=600)

The browser runs in its own LangSmith sandbox, apart from the agent. Defining it adds four tools: browser_snapshot (reads the accessibility tree), browser_run (navigates and acts), browser_screenshot, and browser_handoff.

Scope matters. The docs describe "thread" (the default) as one clean browser per conversation. "user" shares cookies and logins across all of one person's threads. The browser stops after idle_ttl_seconds of no activity (default 600, range 60 to 86,400).

When a site needs a login, CAPTCHA or 2FA code, the agent calls browser_handoff. You get a link in Slack or Studio, take control of the remote browser, sign in, and click Hand back. The password never enters the model's messages or traces. Handoff links expire after 15 minutes. Anyone holding a valid link can drive the browser, so treat it like a credential.

v0.10 also adds async subagents. Set "async": True on a subagent and it runs in the background on its own thread while the main agent keeps talking to the user. One breaking change: connectors.mcp() is gone. MCP servers now go in tools/mcp.py or tools/mcp.ts.

Why it matters for your business

The no-API backlog just got cheaper. Reconciling a portal export, checking order status on a supplier site, filing the same claim form weekly. Those tasks used to need brittle scrapers or a person. A sandboxed agent browser with a human for the login step is a reasonable middle path.

Pick thread scope unless you have a reason not to. LangChain's own docs say user scope shares login state and is not account isolation. An agent in one conversation can affect a session another conversation depends on. For anything that buys or books, keep scopes narrow and approvals on.

It's a beta with no published pricing. Prototype on it. Don't move a revenue-critical workflow onto it until the price and GA terms exist.

Sources: LangChain blog, LangChain docs: Browser use.