# LangChain Managed Deep Agents v0.10 Adds Browser Use and Async Subagents

The public beta update lets agents work through websites and delegate background tasks, with session boundaries and migration details worth checking.

By Paras Madan | 2026-10-30


LangChain released Managed Deep Agents v0.10 in public beta on October 29, adding managed browser sessions and asynchronous subagents. Agents can now interact with websites without an API and continue a conversation while delegated work runs in the background. The update also changes some existing project configuration.

The [release announcement](https://www.langchain.com/blog/managed-deep-agents-browser-use-and-asynchronous-subagents) describes a broader runtime improvement: results and approvals from background work can return to the original Slack thread, and a Stop button can cancel that thread's runs. This is a new version of an existing public beta, rather than the first release of Managed Deep Agents.

## Browser sessions with a human handoff

Developers enable the browser through a `define_browser` declaration at the project root. The agent gets tools for reading the page, interacting with it, taking screenshots and handing control to a person.

The [browser documentation](https://docs.langchain.com/langsmith/python/managed-deep-agents-browser-use) offers two session scopes. Thread scope, the default, keeps a conversation's browser separate from other conversations. User scope shares cookies and logins across one authenticated person's threads, while giving each thread a separate window.

That distinction deserves attention before connecting an account. Separate windows prevent two conversations from clicking the same page, but shared cookies still mean that signing out or switching accounts can affect another thread. The idle timeout stops the browser; it does not log the person out of websites.

For sign-in and other human-only steps, a handoff lets the person interact directly with the website. While they control that window, the agent cannot act in it. Handoff links expire after 15 minutes and need to stay private because anyone holding a valid link can control the browser.

## Why this helps coding agents

LangChain's [technical explanation](https://www.langchain.com/blog/browser-use-in-managed-deep-agents) describes a private connection between the remote browser and the agent's sandbox. A browser can reach an app running there without requiring the developer to publish a public preview URL.

That makes a useful testing loop possible: change a form, start the application, interact with the rendered page and inspect the result. The browser tools use Stagehand and support sequences of actions in one call, reducing how much page content has to enter the model's context.

The practical benefit depends on the test. Seeing a page load is weaker evidence than completing the flow that the change affected. A signup test, for example, should also cover an invalid submission and what happens after the user corrects it.

## Background tasks and migration details

Previously, an inline subagent could keep the orchestrating agent waiting. With an asynchronous subagent, the main conversation can continue while the helper works. Results can wake an idle agent or arrive during an active run; a pending approval remains pending.

The release also removes `connectors.mcp()` and the `connectors/` layout. MCP declarations move to `tools/mcp.ts` or `tools/mcp.py`. The sandbox `scope` option is removed, and shell environment values now take precedence over project `.env` values for `LANGSMITH_*` variables.

**Browser access and background execution increase what an agent can do, and what a developer needs to coordinate.** Before migrating a production workflow, test cancellation, shared account state and late-arriving results. The useful improvement is keeping work responsive without losing track of who authorized it or where its outcome belongs.

I’m Paras. I work at the intersection of AI, engineering and distribution.

I help technical founders explain agent features through clear workflows, implementation details and honest tradeoffs that developers can use.

[Book a 30-minute call](https://cal.com/paras-madan/30min)