Agents that explored widely before tackling hard cases averaged 74.54% on 4 tasks, versus 56.50% when they went straight to the hard cases, so start wide.

AI agents often fumble in unfamiliar apps. The usual fix is collecting new data and retraining, which is expensive.

In RSIAgent, a curriculum agent invents practice tasks, an actor solves them with code, and a separate verifier checks each result. Practice starts broad across related tasks, then goes deep on the real task and its hard cases.

Overall the paper says, Before putting an agent to work in a new app, let it practice broad, then deep, with a separate verifier deciding what's saved.