Meta Engineering 2026-10-23-3
The Thin Orchestrator: How We Made Our AI Agents More Reliable by Giving the Router Less to Do
Meta's orchestrator is 1,300 lines, routes 20+ workflows to 14 subagents and does no domain work; only step 6 of its fixed 7-step protocol involves an LLM. A single agent hit 70-80% success with per-task manual prompts and 10-20% without, and column config prep fell from 3-4 days to under an hour (self-reported). It is the same move as 1Password's, applied to routing: the model gets isolated context to be creative in, and nothing deterministic is left for it to hallucinate.