Your agents can now work through a web app running in their own environment the way a person would: open a page, click, fill in and submit a form, land on the next page, and screenshot each step, staying signed in the whole way.
Read, act, read. The agent asks for the page's interactive elements as an outline, each tagged with a reference such as @e3. It clicks, types, presses or scrolls against those references, then reads the page again to see what happened. A form that lands somewhere unexpected shows up in that second read.
One browser per session. The browser stays open across calls, so cookies, a login and the current page carry from one step to the next. Each session gets its own browser, so two sessions on the same machine never share cookies or a login.
Permanent screenshot links. Every capture comes back with a permanent shareable link that keeps working after the machine is gone. A pull request the agent opens carries the screenshots it took during the task.
Linux only. The browser runs on Linux environments. macOS environments do not have it.
Start here
Ask an agent to open a page and show you what it sees. The Agent's Browser has the details.