LangChain’s October 29, 2026 update adds browser control and background work to Managed Deep Agents. It gives teams building workplace assistants a way to handle website-only tasks and keep conversations moving during longer jobs. Developers can also use the browser to check web applications through their interfaces.
This expands an existing developer platform, not a ready-made office assistant. Teams supply the instructions, tools and AI model; LangChain’s LangSmith service runs the underlying software. The product remains in public beta.
A browser for tasks without a software connection
The browser can navigate, click and fill in forms. But LangChain’s browser guide recommends using a suitable API—a direct software connection—when one exists. For a business choosing what to automate, the useful starting point is the step that currently requires someone to use a website.
For login or two-factor authentication, the assistant can hand the remote browser to a person to sign in on the website, not in chat, then return control. The agent cannot operate that window during the handoff, but sensitive page content may still be visible to it afterwards.
Login sharing is a separate choice. By default, browser state belongs to one conversation and is shared by its authorized participants. An optional user setting carries an authenticated person’s cookies and logins across their conversations. That saves repeated sign-ins, but logging out or switching accounts can affect their other conversations. Ending an agent run does not log the person out.
Background helpers share files, not just results
Previously, a delegated task kept the main agent waiting. The new asynchronous subagents—background helpers with their own conversations—can work while the main agent continues talking to the user. It can check, redirect or cancel their work and receive completion notifications.
The documentation says helpers running as the same user share the parent agent’s sandbox, its working environment and files. Separate conversations therefore do not mean separate copies of a document. Teams delegating concurrent edits need to coordinate which files each helper can change.
Approval handling also depends on the interface. Slack can show a helper’s approval request in the original conversation; other channels require it to be handled in LangSmith Studio or through the API. A completed background task does not override an approval already holding up the main agent.
The immediate gain is a more responsive conversation. These changes do not, by themselves, establish faster task completion or lower costs.
What to check before adopting it
Deployment requires a workspace with Managed Deep Agents beta access. LangSmith’s beta policy says interfaces and behavior may still change, and pricing is not necessarily final. The release announcement gives no date for general availability.
LangSmith lists its Plus plan at US$39 per seat per month plus usage, including one small serverless deployment. Additional deployment and sandbox resources have usage charges. Teams also need model-provider access, so the seat price alone is not a complete budget for an assistant or a flat price per website task.
Existing projects also face a migration. Version 0.10 removes the old MCP connector layout, used to connect agents to external tools. Those declarations must move to tools/mcp.py or tools/mcp.ts. The old sandbox scope setting is removed too; it is distinct from the new browser-sharing setting.
For the related decision of how to build and try reusable workplace assistants, see RohitAI’s coverage of Windmill’s chat-based agent-building workflow.