Improve Agents without Retraining: Parallel Tree Search with Off-Policy Correction