Meta says a thin AI-agent orchestrator cut data-column setup to under an hour

Meta says it has cut the time needed to prepare configuration for a new data column from three to four days by hand—or about a day with its earlier system—to under an hour, using an AI-agent architecture built around a deliberately limited orchestrator. The company’s engineering team says the design routes work and checks reliability while leaving domain-specific decisions to specialist agents.

The system supports data-column onboarding across Meta recommendation systems, a process the company describes as spanning three code repositories, three programming languages and as many as six pipeline stages. Meta says its first approach packed the system’s domain knowledge into a single agent. That setup worked in demos but began producing errors in production, including applying conventions from one subsystem or data source to another.

Meta’s alternative splits the work between a router and 14 specialist agents. The orchestrator classifies requests, gathers required information, runs checks and delegates; it does not generate configurations or make domain decisions, according to the company. The specialists have isolated contexts and handle tasks such as schema management, code generation for different data sources, and validation.

A fixed seven-step protocol governs each request, with deterministic steps for gathering information and checking requirements before an agent takes on the domain work. Meta says it uses workflow templates to sequence the same pool of specialists for different tasks, including column onboarding, event onboarding and pipeline migration. For the migration workflow, the team says it added a template rather than writing new agent code.

Meta also reports that about two-thirds of changes made by the system land without significant human edits. That figure and the onboarding-time comparison are company-reported results; the post does not provide an independent evaluation or a detailed account of how the success measures were calculated.

The engineering team argues that limiting the orchestrator reduces the risk of errors spreading across a workflow while keeping each specialist focused on a narrower body of knowledge. It says knowledge files are loaded when an agent runs and can be updated without changing the orchestrator. Meta plans to extend the approach to additional pipeline domains and is exploring automated ways to identify failure patterns, with human review required for proposed changes.