Agents Can Now Click Through the Web Themselves: LangSmith Hosts a Browser and Passes 90% of 100 WebVoyager Tasks

Agents can now do work on websites that offer no API: read what renders on a page, click, scroll and fill forms, carrying out flows such as booking travel, canceling a subscription or filing a claim the way a person would. Previously an agent could only call an API or fetch the page with curl, which returns the HTML before any JavaScript renders and cannot click anything. On 100 tasks from WebVoyager, a benchmark of everyday tasks on live websites, the toolset passed 90% of tasks for $1.98 in total, the highest pass rate among the seven toolsets tested.

The browser runs in its own LangSmith sandbox separate from the agent, and the agent drives it through four tools, with browser_run doing most of the work by accepting a whole chain of steps in one call. browser_snapshot reads the page's accessibility tree and labels interactive elements with IDs, browser_run acts through batched actions such as click and fill or a Playwright-style page object, browser_screenshot captures the rendered page, and browser_handoff passes the browser to a person. The tools are built on Stagehand, Browserbase's open source SDK for browser agents, and were tuned by running browser benchmarks, reading traces in LangSmith and validating each change with evals.

At a login page the agent does not receive a password; it calls browser_handoff with a reason, the user signs in on an interactive card in LangSmith Studio and clicks Hand back, and the agent continues already signed in. Giving the agent a password would put it into the model's messages, traces and the model provider, whereas handoff keeps credentials with the user. Handoff ships with every managed browser and works automatically in Managed Deep Agents channels such as Slack; under user scope the sign-in carries over to that user's later threads.

The demonstrated scope is browser use inside LangSmith managed agents: thread- or user-scoped sessions, billing that stops after an idle timeout, login handoff, and a localhost bridge that lets the browser reach an app running in the agent's sandbox for end-to-end testing.

The pass rate and cost come from one benchmark, 100 WebVoyager tasks, and no other benchmark is reported; the testing and tuning details behind the tools are deferred to a later deep dive, with only the benchmark, trace and eval loop described here.