LangChain's Managed Deep Agents v0.10 adds browser use, login handoffs and background subagents

LangChain has shipped version 0.10 of Managed Deep Agents, its hosted service for running agents built on the Deep Agents framework inside LangSmith. Announced on October 29 and available in public beta, the update centres on two additions: a browser the agent can drive, and subagents that work in the background.

A browser for sites without an API

Developers enable the browser by declaring it in a file at the project root. It runs in its own LangSmith sandbox, not on the user's machine, and gives the agent four tools: one that reads the page's accessibility tree and labels clickable elements, one that performs actions, one that takes screenshots, and one that hands control to a person. According to a companion post, the tools build on Stagehand, an open source SDK from Browserbase.

LangChain pitches the feature for tasks such as booking travel or filing a claim on sites with no API, and advises using an API when one exists. It also describes end-to-end app testing, in which the agent starts an app in its sandbox and opens it in the browser. By default each conversation gets a fresh browser; a user-level scope lets sign-ins carry across one person's threads.

When the agent hits a login page, it can call the handoff tool. The person receives a link to the remote browser, signs in, then clicks "Hand back". LangChain says this keeps credentials out of model messages and traces. Handoff links expire after 15 minutes. Slack and LangSmith Studio display the handoff natively; custom front ends must render it themselves.

Subagents that run in the background

Setting an async flag on a subagent makes it run on a separate thread while the user keeps talking to the main agent, which can start, check, update or cancel those tasks. Results are delivered depending on the agent's state: before its next model call if busy, as a new run if idle, or after the user answers a pending approval.

Slack controls and caveats

Slack threads now get a Stop button that cancels running work, and experimental Slack channel handlers control what each event triggers. The release has breaking changes: MCP servers must now be declared in a tools file instead of the removed connectors layout, and the sandbox scope option is gone. LangChain cites an internal benchmark of 100 WebVoyager tasks in which its toolset passed 90%; that figure has not been independently verified. Neither post lists pricing.