OzWizards

DriftPA trains LLM agents to be reliable personal executive assistants in a world that breaks without warning. Real AI assistants fail in production because APIs change schema mid-session, tasks expire while the agent deliberates, and wrong bookings or emails can't be undone. DriftPA is an RL environment that simulates exactly these four failure modes simultaneously: schema drift (field names change mid-episode), time pressure (urgent tasks expire), irreversible actions (book/reply/cancel can't be undone), and policy drift (cancellation windows tighten). An untrained agent scores -9.55 mean reward — it uses stale API fields, misses the boss's urgent email, double-books dinner, and triggers cascade failures. A GRPO-trained agent learns to call list_tools() after drift, prioritize by urgency, and commit to irreversible actions only when safe. Optimal score: +22. Untrained mean: -9.55. That 31-point gap is the problem we solve.

Team