Agents learn to click through websites, handing the browser back to you at login so passwords stay out of traces

Agents can now complete tasks such as booking travel, canceling a subscription or filing a claim on sites that have no API and were built for people clicking through pages. Previously agents could reason through a task but could not take action on websites without an API, and fetching a page with curl produced HTML before JavaScript rendered it, leaving nothing to click. The release notes place this alongside demand from personal and coding agents, and report that DataDome agentic browser requests jumped 4.7x in September alone.

The browser runs in its own LangSmith sandbox, and Managed Deep Agents automatically adds four browser tools to the agent. The tools are browser_snapshot, which reads the page's accessibility tree and assigns element IDs; browser_run, which navigates and interacts through JavaScript or batches of actions against snapshot IDs; browser_screenshot, which captures the page and can save it to the sandbox; and browser_handoff, which gives a person control of the window. By default each conversation gets its own browser, reused across turns and starting with no cookies or logins from other threads; with scope="user", each signed-in person gets one browser whose logins carry across all their conversations.

At a login wall or two-factor prompt, the agent calls browser_handoff to give the window to the user, so the password never enters the model's messages or traces. Most browser agents stop at a login wall or a two-factor prompt, whereas handoff lets the user sign in and the agent then checks the page to see whether the login worked and continues. Handoff returns immediately so the agent can do other work while it waits, and handoff links expire after 15 minutes.

Subagents can run asynchronously, letting the orchestrator agent keep talking to the user while background tasks run. Previously subagents ran inline, so the agent called one and waited for its answer, making the person wait on long tasks and blocking further interaction with the orchestrator. Adding async to a subagent definition enables it; when a subagent finishes, a busy agent checks unfinished tasks before each model call, an idle agent is woken with a new run, and an agent paused on an approval is left alone until the user answers.

The release notes address developers building agents with Managed Deep Agents and describe the browser use and async subagent capabilities of v0.10, available in public beta, together with how they are configured.

The release notes give no success rates or comparative evaluation for browser tasks or async subagents; v0.10 is in public beta and includes breaking changes such as the removal of connectors.mcp() and the sandbox scope option.