Meta describes how an AI agent can be designed to capture the logic and expertise of domain experts, rather than simply storing documents or retrieving relevant information. The system, dubbed an "organizational second brain", was built for a specialized compliance domain, but Meta argues the architecture generalizes to areas like security, finance, engineering, and procurement.
We've built an AI agent that acts as a secondary expert for a given domain, making deep specialist knowledge readily available and preserved for anyone in an organization to access, share, and build upon.
According to Meta, their approach differs from a traditional domain-specific agent by combining a "structured, auditable knowledge architecture" with a self-improvement loop that "compiles expert feedback into verified, regression-tested updates without model retraining". The architecture consists of four layers: a knowledge system to consolidate all institutional knowledge; a reasoning pipeline that separates what the agent knows from how it reasons; an evaluation framework to provide automated benchmarks; and the already mentioned self-improvement loop.
The knowledge system uses expert knowledge distilled from over 200 structured files organized into a strict taxonomy. These include position files, capturing authoritative know-how; taxonomy and vocabulary files, acting as a glossary for entity types, activity categories, classification tiers, and other terms used across the organization; routing indexes, mapping inputs to relevant files without relying solely on embedding similarity; and gateway files, defining tests that determine when the agent can safely apply specialized knowledge.
The reasoning layer uses "recipes" to define how the agent should analyze a problem. Recipes are composable and specify what the agent should examine first, which knowledge to load at each step, which decision procedure to follow, and how to determine when the analysis is complete. Recipes also make it possible to determine whether a failure came from missing knowledge or a flawed reasoning procedure.
A key feature of Meta's approach is the use of predefined checkpoints in the analysis where human control is required and ambiguous cases are escalated to domain experts. When an expert identifies a mistake, the correction is permanently recorded, whether it stems from missing knowledge, a flawed reasoning procedure, or genuine ambiguity among experts.
The key idea that makes this self-improvement flywheel possible is to keep institutional knowledge out of the model's weights and instead store it in version-controlled text files that are accessed by explicit reasoning procedures and rigorously tested:
Expert corrections are diagnosed to their root cause, compiled into minimal verified edits, and evaluated against replay and regression tests before they are reviewed and landed. Each fix is then folded back into the regression suite, so the gain is permanent.
Based on Meta's own assessment, its approach to building organizational experts with agents reduced the time required for individual assessments from days to minutes, with domain experts judging the results useful almost all the time. Furthermore, the system automated knowledge improvements that previously required engineering sprints and experienced zero regressions across improvement cycles.
There is much more to Meta's approach to creating agents capable of serving as organizational experts than can be covered here. Be sure to read the original article for the full details.