Actors that browse the web

Actors can now open a real browser, read pages and take actions, grounded in screenshots of what they actually see.

• Browser automationNew

  Actors run in a hosted browser and act on live pages from plain-language instructions, so they work through real products rather than talk about them.

• Screenshot-grounded actionsNew

  Every step starts from a screenshot of the current page, so an actor sees what you would see before it clicks, types or navigates.

• Multi-step tool loopImproved

  Actors chain actions across a session, reading the result of each step before choosing the next, which makes longer tasks far more reliable.