The Thin Orchestrator: How We Made Our AI Agents More Reliable by Giving the Router Less to Do

Why it matters

Meta reports a lesson in AI agent architecture drawn from automating data column onboarding for its recommendation systems. The company initially built a single agent loaded with all domain knowledge, which performed well in demonstrations but hallucinated once deployed. Meta found that beyond a certain context threshold, additional knowledge degraded performance, as the agent mixed up conventions across subsystems and applied code generation patterns to the wrong data sources. The remedy was architectural rather than prompt-based: a deliberately thin orchestrator that only routes work, gathers fields, and enforces reliability checks, while specialized subagents handle domain reasoning in isolated contexts. Workflow templates, described as deterministic state machines, sequence the same pool of specialists for different use cases, so new capabilities required no new agent code. Meta also cites a Mount Sinai preprint on multiagent systems maintaining accuracy where single agents degrade, attributing the gap to task isolation. The piece is a first-party account of context interference and how separation of concerns addressed it in production.