Conduct a structured multi‑agent debate with fact‑checking
RL env to classify, route and reply to emails