MiniMax M2.7 lands with native agent teams

This is the stronger bet than trying to fake autonomy with a giant system prompt. If the claims survive external replication, M2.7 looks less like a benchmark trophy and more like a blueprint for stable tool-using systems.

• –MiniMax says Agent Teams need role boundaries, adversarial reasoning, protocol adherence, and behavioral differentiation, which is exactly the stuff prompt-only agents tend to lose first.
• –The self-evolution loop is the real novelty: M2.7 reportedly updated memory, built dozens of skills, and iterated on its own scaffold for 100+ rounds.
• –MiniMax also claims M2.7 can cover 30%-50% of the RL team's workflow and posts strong scores on SWE-Pro, GDPval-AA, and Terminal Bench 2, so this is meant to be production infrastructure, not a toy.
• –The caveat is straightforward: these are vendor-reported numbers, so third-party evals will decide whether the architecture generalizes beyond MiniMax's own harness.