On October 29, 2026, LangChain released Managed Deep Agents v0.10 in public beta. Its headline change looks modest but matters: the managed agent gets its own browser. The reason is a gap every API strategy leaves open — many websites are built for people who click, fill in forms and read pages, not for clean interfaces. Where no API exists, an agent could previously only watch.
What the browser gives the agent
The browser runs remotely in the LangSmith sandbox, separate from the agent, not on the user's machine. Developers declare it, and the agent automatically gains four tools.
| Tool | Job |
|---|---|
| browser_snapshot | Reads the page's accessibility tree and assigns element IDs |
| browser_run | Navigates and interacts, through JavaScript or batches of actions against snapshot IDs |
| browser_screenshot | Captures the page as an image, optionally saved |
| browser_handoff | Hands the window to a person, for sign-in or two-factor checks |
The key trick is batching: browser_run can open, fill, click and check in a single call instead of paying a model request per action. That keeps long tasks faster and keeps less page content in the model's working context. By default, each conversation gets its own browser, reused across turns and starting with no cookies or logins from other threads. A different configuration gives each user one browser whose logins carry across all their conversations, so they sign in only once.
Why sign-in stays a handoff
When the agent hits a login wall or a two-factor prompt, it calls browser_handoff and explains why it needs help. The person signs in through an interactive browser card in LangSmith Studio and hands control back; the agent then continues in the authenticated browser. Handoff links expire after 15 minutes, and the same handoff works in channels such as Slack. This is a deliberate boundary: credentials and checks stay in human hands instead of the agent guessing or storing passwords.
The second change concerns long-running work: asynchronous subagents, plus a new runtime.thread record of where each thread started and who is on it. Combined with the private connection for app testing — the agent starts the app in its own sandbox and directs the browser at it through localhost, with no public URL required — the use case is clear: complete web tasks that lack an API, and test applications by clicking through them as a user would. As a first performance claim, LangChain cites its own tests: 90% of 100 WebVoyager tasks on live websites for $1.98 in total. That is a vendor number, and LangChain promises a deeper account of testing and tuning next week. Anyone sizing this up should wait for independent replication before turning browser agents loose on business-critical flows.
