Meta's Recipe for Building Agents as "Organizational Second Brains"

Details

Meta has detailed the design of what it calls an "organizational second brain": an AI agent architecture built to preserve and apply deep specialist knowledge at scale. The system was originally deployed for a compliance domain and is now described as generalizable to security, finance, engineering, and procurement. The architecture has four layers. A knowledge system consolidates expertise from over 200 structured files organized in strict taxonomy, including position files (authoritative know-how), vocabulary files, routing indexes, and gateway files that gate when specialized knowledge applies. A reasoning layer uses composable "recipes" — step-by-step analysis blueprints that specify which knowledge to load and when, making failures attributable to either missing knowledge or flawed reasoning. An evaluation framework provides automated benchmarks. A self-improvement loop permanently records corrections from human experts and compiles them into regression-tested knowledge updates — without retraining the underlying model. A notable design choice: the system includes predefined checkpoints where human control is required, with ambiguous cases escalated to domain experts. This makes the agent auditable and correctable in production. The separation of the reasoning procedure from the knowledge base is particularly useful for diagnosing errors in complex compliance-type workflows. The write-up is light on quantitative results and heavy on architectural description, so the claimed benefits — reduced hallucination, better auditability, lower retraining costs — are not independently verified here.