The development
What happened
Meta’s 19 October engineering report describes a deterministic router that delegates data-pipeline work to isolated specialist AI agents using large language models for domain reasoning. The system handles data-column onboarding across three repositories and three languages.
Key facts
A roughly 1,300-line orchestrator routes more than 20 workflows to 14 specialists, without generating domain code or configurations itself.
Fixed checks gather inputs, detect duplicates and enforce reliability gates before delegation; results and errors are checked afterward.
Workflow templates reuse specialists in different sequences, with domain knowledge loaded when needed.
Meta reports preparation falling from three–four manual days, or about one day with its earlier configuration system, to under an hour. About two-thirds of changes need no significant human edits.
The context
Why it matters
The design makes routing and verification explicit while limiting context interference. The efficiency figures are company-reported observations, not independent controlled evidence that additional agents improve every workflow.