Sandworld in two minutes.

What you'll see

1. The worlds: working twins of a grocery app, a wallet, an airline, invoicing, an issue tracker and email, plus one company using three apps together.
2. One world up close: its real screen, a task and the tools an agent gets.
3. Two real agent runs, and the verifier's checks that pass or fail each one.
4. Published model results on the Arena, and why each failed episode failed.
5. The open datasets: verified trajectories and preference pairs, with a held-out split.
6. A task done by hand in the grocery twin, then verified.
7. The cross-app world, and the mobile twins running as an Android app.