Stacks/Data / ML platform
Meta's thin orchestrator for data pipeline agents
Meta's recommendation-system data teams replaced a single 'smart' agent that hallucinated in production with a thin deterministic router and 14 isolated specialist subagents. The orchestrator never reasons about code; it classifies intent, gathers fields, runs reliability checks, and delegates. New use cases are added as workflow templates, not new agents.
- Use case
- Automating data column onboarding across Meta recommendation-system pipelines, spanning three code repositories, three programming languages, and up to six pipeline stages
- Category
- Data / ML platform
- Team size
- not stated
- Orchestrator / framework
- A deterministic 1,300-line thin orchestrator that performs zero domain work. It routes 20+ workflows to 14 specialized subagents using a fixed 7-step protocol; only the delegated step involves LLM reasoning. Workflow templates (deterministic state machines) reuse the same pool of subagents in different sequences for column onboarding, event onboarding, and pipeline migration.
- Memory
- 25+ knowledge files, organized by workflow domain, are loaded on demand at subagent execution time rather than embedded in the orchestrator. Updates take effect on the next session without an orchestrator change.
- Evals
- A single-agent alternative reached 70–80% success only with manual prompt engineering per task, and 10–20% without it. About two-thirds of the thin-orchestrator system's changes land without significant human edits. Configuration prep that took three to four days by hand, or about a day with an earlier config-driven system, now finishes in under an hour.
- Observability
- not stated
- Monthly cost
- not stated
- Models
- not stated
- Observability tools
- not stated
- Other tools
- Config-as-code (Python configurations driving C++ code generation)
- Source
- https://engineering.fb.com/2026/10/19/ml-applications/thin-orchestrator-ai-agent-architecture/
Plawan Rath, Akhil Jindal, and Ashwini Verma · Oct 18, 2026