Meta built a thin orchestrator for its data pipeline agents. It routes work and enforces checks but does no domain reasoning. Fourteen specialized subagents handle domain tasks in isolation. A 7-step protocol keeps routing deterministic. Workflow templates sequence existing agents for new use cases. Configuration prep for a new column now takes under an hour. Two-thirds of changes land without human edits. The system scales by adding templates and knowledge files. Failures are contained within subagents. Meta shared the pattern for others to adopt.
What changed
A single AI agent carried all domain knowledge and made all decisions.
What it unlocks
Running multi-step workflows across heterogeneous systems without cross-contamination or hallucination.
- 1,300-line orchestrator
- 20+ workflows
- 14 specialized agents
- 7-step protocol
What you need to act on it
- a multi-agent architecture
- deterministic orchestrator layer
- specialized subagents with isolated context
Sources