Simulating Environments with Reasoning Models for Agent Training