About
AI agents will do real work in real apps. They should practise somewhere safe first.
Why
Agents are graded by verifiers. Most verifiers are never tested.
The fastest way to improve an agent is to let it act, check the result and learn from it. That loop is only as good as the check. A grader that passes a wrong answer teaches the wrong lesson at scale.
So every Sandworld task ships with a verifier that inspects the final state of the app, and every verifier is attacked before it ships: agents that do nothing, act at random, stop early or make changes nobody asked for must fail — and the report is public.
Today the catalog has 6 app twins and 1 cross-app world, with 117 tasks. We build custom twins for labs and pair them with expert demonstrations recorded inside the same worlds.
Founder
Revanth Rajeev
Founder, Sandworld
Revanth builds and ships apps, including Android apps on Google Play, and built Sandworld's app twins, task verifiers and expert platform.
He started Sandworld because AI agents are about to do real work in the apps people rely on, and the teams training them need a safe place to practise and a grader they can trust. If you run evaluations or post-training at a lab, he'd like to hear what your agents get wrong.