RL environment for enterprise customer-support agents that trains with zero human annotation. A Dialogue Agent and an adversarial, GRPO-trained Customer Agent co-evolve in a multi-turn banking workspace, where the Customer Agent conceals intent to induce misroutes and the Dialogue Agent retrains on its own failure transcripts to route calls correctly. Tasks span transactions, disputes, card issues, transfers, and fraud escalations.