🧭Meta's Thin Orchestrator Hits 65.3% Accuracy Across 80-Task Agent Batches

Stop stuffing your AI orchestrator with domain knowledge

The conventional move is making the orchestrator the smartest part of your agent system. Meta went the other way. Its data pipeline agents use a thin router that owns sequencing and reliability checks and hands every domain decision to specialists. Honestly? That's the more interesting design. If you're building production agents, this is a direct critique of the stuff-everything-into-one-prompt habit. A single agent can cross-contaminate concepts and degrade fast as the workload grows. The thin orchestrator keeps context isolated, so a subagent working on one domain never sees the noise from another. Knowledge files load on demand at execution time, not baked into the orchestrator. The numbers: a preprint from the Windreich Department of Artificial Intelligence and Human Health reports orchestrated multiagent systems held 65.3% accuracy at 80-task batches. The system runs a fixed 7-step protocol on every request and maintains 25+ knowledge files organized by workflow domain. Per Meta's account, prep for a new column takes under an hour. The catch is that the preprint is early, so treat the accuracy figure as a signal, not a benchmark you can bank on.