Self-play is most useful for narrow domains like competition and social negotiation, but adversarial setups more broadly (debate, prover-verifier, LLM-as-judge) can create diversity by incentivizing agents to differentiate their approaches from each other.