Meta improves AI agent reliability with thin orchestrator pattern

• Meta's thin orchestrator architecture routes 20+ workflows to 14 specialized agents, reducing data column onboarding time from 3-4 days to under an hour.
• The system achieves about two-thirds of changes without significant human edits by isolating domain reasoning in subagents and keeping the central router deterministic.
• A 1,300-line orchestrator handles sequencing and reliability checks while 14 subagents manage specific tasks like schema management, code generation, and validation.
• Workflow templates allow new use cases to be added by sequencing existing agents differently, requiring zero new agent code.

Architecture Design

• The orchestrator follows a fixed 7-step protocol where only the delegation step involves LLM reasoning within an isolated subagent.
• Subagents load knowledge files on demand at runtime, allowing updates to take effect immediately without redeployment.
• The design separates deterministic routing decisions from probabilistic domain reasoning to prevent context cross-contamination.

Implementation Benefits

• Context isolation prevents agents from applying patterns from one data source to another, reducing hallucinations.
• Failures are contained within individual subagents, preventing corruption of the orchestrator's state or other agents' contexts.
• Adding new capabilities is templated work, as new subagents follow a standardized structure with knowledge sources, checklists, and troubleshooting mappings.