The Pulse
LangChain Adds Browser Control and Async Subagents to v0.10
Managed Deep Agents v0.10 adds browser tools for websites without APIs and lets subagents handle tasks in the background while the main agent continues talking with users. The public beta also changes Slack run controls and removes some pro
AI.info Team · Written with AI · fact-checked ·
LangChain’s Managed Deep Agents v0.10, announced October 29, 2026, gives agents a browser for websites without APIs and lets delegated subagents work in the background. The public beta pairs those features with changes to Slack handling, run controls and project configuration.
The browser runs in its own LangSmith sandbox and comes with four tools: one reads a page’s accessibility tree, another navigates and interacts with it, a third takes screenshots, and a fourth hands control of the window to a person. LangChain says the tools are intended for tasks such as booking travel, canceling subscriptions and filing claims on sites designed for people rather than API calls.
Browser sessions keep logins separate
By default, each conversation gets its own browser session, reused across turns but starting without cookies or logins from other threads. Developers can instead set the browser scope to a user, allowing that person’s logins to persist across their conversations. That makes the choice a trade-off between isolated sessions and convenience for repeat use.
When an agent reaches a login or two-factor prompt, it can call browser_handoff to give the user control of the remote window. The user signs in and hands control back; LangChain says the password stays out of the model’s messages and traces. Handoff links expire after 15 minutes, and the agent can continue other work while waiting for the user.
Subagents no longer have to hold the conversation
Before v0.10, Managed Deep Agents subagents ran inline: the main agent waited for the helper’s answer before continuing. Developers can now mark a subagent definition as asynchronous, sending its task to a separate thread while the orchestrator keeps responding to the user.
The orchestrator can start, check, update or cancel those tasks, and the subagents notify it when they finish rather than requiring a polling loop. If the main agent is busy, it checks for completed tasks before its next model call; if idle, a completion starts a new run. When an agent is paused for approval, the result waits until the user responds. Async subagents also share the agent’s sandbox, including files already written there.
Slack threads retain their point of origin
Version 0.10 adds runtime.thread, which records where a conversation began and who is participating. LangChain says that lets scheduled runs, Studio resumptions and subagent wake-ups send replies to the original Slack thread. Slack threads also get a Stop button to cancel runs, while Studio can cancel runs on Slack threads.
The release includes experimental Slack event handlers and a bring-your-own-app option during deployment. It also changes project configuration: connectors.mcp() and the connectors/ layout have been removed, and MCP servers must now be declared in tools/mcp.ts or tools/mcp.py. LangChain says shell environment variables take precedence over a project’s .env file for LANGSMITH_* variables. The release also removes the sandbox scope option.
Public beta, with configuration changes
LangChain released Managed Deep Agents v0.10 as a public beta. Developers can start a project with the mda init command from the managed-deepagents package and deploy it with mda deploy, according to the company’s instructions.
The release adds browser use and background delegation to Managed Deep Agents, while removing connector configuration and the sandbox scope option. Existing projects that use the removed settings will need to account for those changes. The browser handoff flow lets a person sign in; the async task tools let the orchestrator manage delegated work.