Meta open-sources thin orchestrator pattern for multi-agent AI systems

Meta Engineering published a deep dive into its 'thin orchestrator' architecture for production AI agents. The system replaces a single knowledge-heavy agent that hallucinated in production with a minimal router and 14 domain-specialist subagents. The orchestrator handles only intent classification, field gathering, reliability checks, and delegation; all domain reasoning happens in isolated subagents with restricted context. The company reports that data column onboarding time dropped from three to four days manually, or about a day with an earlier system, to under an hour. About two-thirds of changes now land without significant human edits.

How the thin orchestrator works

• 1,300-line deterministic orchestrator routes 20+ workflows to 14 specialized subagents; performs zero domain work itself
• 7-step fixed protocol: read instruction file, check for duplicates, gather fields, auto-discover inferred values, reliability gate check, delegate to subagent, handle results
• Only step 6 uses LLM reasoning, and it runs inside an isolated subagent with domain-specific context loaded on demand from knowledge files (typically a few thousand tokens each)
• Workflow templates are deterministic state machines defining phase ordering; new use cases (e.g., pipeline migration) require only new templates, not new agents
• Context isolation prevents cross-contamination: the population agent for data source A never sees data source B's patterns
• 25+ knowledge files maintained by domain; updates take effect immediately without orchestrator changes

Why it matters

The pattern offers a concrete alternative to the common 'bigger context window' approach for agent reliability. Meta's explicit open publication of the architecture, including failure modes of single-agent systems and quantitative results, gives engineering teams a reference implementation for production multi-agent systems.