LangChain AI Updates: October 30, 2026

1. Managed Deep Agents Get a Remote Browser With Sign-In Handoff

LangChain added browser use to Managed Deep Agents, aimed at web tasks with no API (such as booking a reservation) and at end-to-end testing of app changes before merge. Developers export a browser from browser.py with define_browser(scope="thread", idle_ttl_seconds=600); thread scope gives each conversation a fresh browser, while user scope shares cookies and logins across a signed-in user’s threads, and idle browsers shut down after 10 minutes by default. The browser runs in its own LangSmith sandbox, connected to the agent’s sandbox by a private link that provides a localhost bridge, so an app started on, for example, port 5173 can be tested without a public URL. Four tools are added: browser_snapshot (accessibility tree with element IDs), browser_run (batched actions or Playwright-style JavaScript, translated by Browserbase’s open-source Stagehand), browser_screenshot and browser_handoff. Because browser_run accepts code, the agent can chain open, fill, click and check steps in one call and return only the results it needs, keeping most of the page out of context. When the agent hits a login page, browser_handoff gives the user the live browser in Studio or Slack, so passwords stay out of model messages, traces and model providers. LangChain reports a 90% pass rate on 100 WebVoyager tasks for $1.98 total, the best of seven toolsets tested, with Browserbase’s MCP server next at 89% and 9x the cost. Passing tests can be saved as Playwright-style files that replay without model calls until the page changes. Open-source Deep Agents users can bring their own browser provider. Source

2. Managed Deep Agents v0.10 Adds Async Subagents and Slack Controls

LangChain shipped Managed Deep Agents v0.10 in public beta. Setting "async": True on a subagent in define_deep_agent runs it on its own thread while the main agent keeps talking to the user, with tools to start, check, update and cancel tasks, and every served agent gets a background copy of the general-purpose subagent. The agent is notified on completion instead of polling: a busy agent picks up results before its next model call, an idle agent is woken with a new run, and a pending approval is left alone until the user answers. Async subagents share the agent’s sandbox, appear under its run in LangSmith, and post approvals to the original Slack thread. A new runtime.thread records where each thread started and who is on it, so scheduled runs and subagent wake-ups reply in the right Slack thread. Slack threads gain a Stop button, experimental channel handlers (onEvent, status, reply: "none"), a bring-your-own Slack app option in mda deploy, and resolveUser for mapping a verified Slack sender to a user in the developer’s system. Breaking changes remove connectors.mcp() and the connectors/ layout in favor of tools/mcp.ts or tools/mcp.py, drop the sandbox scope option, and give shell LANGSMITH_* variables precedence over a project .env. Source