A graphical user interface agent, meaning software that operates a website on your behalf, needs somewhere to be trained and tested. Live sites change, block access and hide pages behind logins; the stand-in sites built for training are often shallow and unlike the real thing.
The team built replicas of real websites that run in containers and behave the same every time, growing the set task by task: as new tasks demanded features, added them while people checked visuals, behaviour and task quality.
It holds 3,518 tasks, many behind a login. An agent with four billion parameters, trained by , meaning it learned by trial and error inside these replicas, handled tasks it had not seen and matched leading open agents on public tests.