LangChain's managed agents can now work on sites with no API and hand you the browser to sign in, so passwords stay out of the model

The hardest tasks to automate are often the ones stuck behind a login page with no API.

What happened

LangChain released Managed Deep Agents v0.10 in public beta. The main addition is browser use: the agent gets a browser that runs in its own LangSmith sandbox, with tools to read the page, click and type, take screenshots, and hand control to a person.

That handoff is the interesting part. When a site asks for a login or a two factor code, the agent hands the window to you, you sign in, and you hand it back. LangChain says this keeps your password out of the model's messages and traces. Handoff links expire after 15 minutes. By default each conversation gets its own browser with no cookies from other threads. A user scope keeps one person's logins across their conversations, so they sign in once.

The release also adds async subagents that run long jobs in the background and report back when done, replies that return to the Slack thread a request started in, and a Stop button in Slack threads. There are breaking changes too: the old MCP connectors layout was removed.

My take

Most operations teams still have a few portals with no API: supplier sites, insurance and government portals, old booking systems. Today that work goes to a VA or a fragile script.

How I would approach browser agents for those:

1. Use the API when one exists. A browser is the last resort, not the default.
2. Start read only. Pulling a report or checking a status is a safer first pilot than submitting forms.
3. Keep logins human. The handoff pattern is the right one. Do not paste client passwords into prompts.
4. Pick the scope on purpose. A shared company account and a personal account need different browser sessions.
5. Keep screenshots. They are your audit trail when someone asks what the agent did.

It is beta, so pilot it on one portal before you promise anything to a client.