The consensus architecture for multi-agent systems in 2026 is: orchestrator + isolated subagents . A single coordinator holds context, spawns specialists, merges results. You've probably built something like this. A coordinator agent fans out to a research subagent, a code-generation subagent, maybe a QA subagent. The orchestrator waits, collects, synthesizes. It works. But there's a hidden assumption baked into most implementations: agents communicate through you. The coordinator is the hub. Every message routes through your application code. Subagents don't know about each other. When they need data from a peer, they go back up the stack to the orchestrator, which fetches it, hands it down. That's fine for small fleets. It breaks at scale. What "Agents Communicate Through You" Costs When you're the message bus for your agent fleet, every inter-agent exchange adds a round trip through your application. At 5 agents, invisible. At 50, painful. At 500, you've built a distributed system bottleneck.…