The Thin Orchestrator: How We Made Our AI Agents More Reliable by Giving the Router Less to Do
Meta improved production AI reliability by keeping orchestrators thin and routing to specialized agents
“We built a 1,300-line orchestrator that routes 20+ workflows to 14 specialized AI agents. It performs zero domain work itself. That's the point.”
Meta's engineering team found that loading a single AI agent with all domain knowledge caused hallucinations in production due to context overload and concept cross-contamination. Their solution—the 'thin orchestrator' pattern—separates routing and reliability checks from domain reasoning, delegating the latter to 14 specialized agents. A referenced Mount Sinai preprint corroborates the approach, showing multiagent systems maintained 65.3% accuracy at scale where single agents dropped to 16.6%.