From the source
Meta says it improved the reliability and efficiency of its data pipeline agents by keeping the orchestration layer deliberately thin, letting a router sequence work while 14 specialized agents handle domain reasoning.
- The company built a 1,300-line orchestrator that routes more than 20 workflows to those agents and performs no domain work itself.
- Meta's first attempt was a single agent loaded with all domain knowledge, configuration schemas, code generation rules and population patterns for more than 10 data sources; it worked in demos but hallucinated in production, cross-contaminating concepts between subsystems.
- Meta cites a Mount Sinai preprint finding orchestrated multiagent systems held 65.3% accuracy at 80-task batches where single agents fell to 16.6%, while reducing token usage up to 65-fold.
- An alternative single-agent approach reached a 70-80% success rate only with manual prompt engineering per task, dropping to 10-20% without it.
- The orchestrator follows a fixed seven-step protocol, of which only delegation to a subagent involves LLM reasoning, and workflow templates sequence the same 14 subagents for column onboarding, event onboarding and pipeline migration.
- Knowledge files are loaded on demand at subagent execution time, with 25+ files maintained by workflow domain.
