Five major vendors quietly killed peer-to-peer agent chat. If yours still does it, you're behind.

The 2024 idea was that more agents talking to each other equals more intelligence. GroupChat. Everyone wired their agents to message each other. That pattern just lost, and it lost decisively.

Anthropic, OpenAI, AutoGen, Cognition, and LangChain independently settled on the same default: one orchestrator that owns full conversation context, spawning ephemeral isolated subagents that return compressed summaries with no peer-to-peer chatter. Five vendors, no coordination, same conclusion. When that happens it's usually because the alternative stopped working in production. Latent Space is calling the broader moment "Meta-Harness Summer", the shift from building individual harnesses to building harnesses that compose other harnesses.

There's math underneath the convergence now too. A new study on the Co-Failure Ceiling tested 67 frontier models and proved multi-model systems (routing, voting, cascades, mixture-of-agents) can never exceed accuracy of 1 minus beta, where beta is the rate at which every model is wrong on the same query. The usual diagnostic, average pairwise error correlation, can't even identify beta, which means a lot of reported ensemble gains are chasing improvements that are mathematically unreachable. Peer-to-peer chatter doesn't beat that ceiling. It just spends more tokens hitting it.

The production survivors back this up. Roughly 40% of multi-agent pilots collapse within six months, and the post-mortems show every surviving system uses a structured "P2" contract for orchestrator-to-subagent handoffs: explicit objective, required output format, tool guidance, hard task boundaries, dedicated system prompt. Loose "go figure it out" delegation is the failure mode.

What to do: if your agents message each other directly, rip it out. Move to one orchestrator with a near-empty context that fans out to isolated subagents, each handed a typed request like an API call, each returning a 1,000-2,000 token condensed summary. Measure your beta before you invest in any orchestration layer. It's one number, and it tells you whether the layer can possibly pay off.