Training standard AI models against a diverse pool of opponents — rather than building complex hardcoded coordination rules — is enough to produce cooperative multi-agent systems that adapt to each ...