A 20B search agent achieves 0.730 average curated recall across eight benchmarks by training RL on explicit state rather than integrating state management into the policy.