Meta Uses Thin Orchestrator to Improve AI Agent Reliability

Meta Platforms detailed in an engineering blog post how it improved the reliability and efficiency of data pipeline agents by limiting a central orchestrator to routing and reliability checks. Domain reasoning is assigned to specialist agents with isolated context.

The orchestrator contains 1,300 lines of code and routes more than 20 workflows to 14 specialist agents. It uses a fixed protocol with seven steps to check for duplicates, collect required information, identify warnings, delegate work and handle results. Only the delegation step uses large language model reasoning.

Meta's first approach was a single agent loaded with all the domain knowledge, which began cross-contaminating concepts in production. An alternative single-agent approach it evaluated succeeded 70% to 80% of the time with a custom prompt for each task and 10% to 20% without one. The specialist architecture reduced configuration preparation for a new data column from three to four days by hand to under an hour, while about two thirds of changes required no significant human edits.