Meta describes an "organizational second brain" that encodes how its experts reason

Meta's engineering team published an account of how it encoded institutional expertise into an AI agent for a specific compliance domain. The premise: in large organizations the most valuable specialist knowledge is never written down. In compliance work the same questions recur across hundreds of product reviews, expert assessments take days of manual research, and inconsistency between assessments creates organizational risk.

The agent rests on three parts: a knowledge system acting as the organization's "second brain", a reasoning layer that mirrors how domain experts actually think, and an automated improvement pipeline. The design has four layers, and according to the post they are mutually dependent, removing one degrades the rest.

Knowledge distilled in advance

Meta argues that an agent retrieving document chunks at inference time must re-derive expert reasoning on every run, which is slow, error-prone and inconsistent. Instead, a long-running offline process reasons through source documents and distills them into structured knowledge files: curated statements of how the organization interprets its domain, with constraints, boundaries and routing implications made machine-readable.

More than 200 files sit in a strict taxonomy. Each declares its dependencies (depends_on) and consumers (referenced_by) in YAML frontmatter, forming a bidirectional dependency graph so that when one file changes, the affected files can be traced. That matters because a self-improvement loop proposes automated edits. The same loop lets the agent absorb feedback from human experts without retraining the underlying model.

Knowledge is partitioned between the curated wiki and supplementary retrieval (RAG) by information density and expected usage frequency. Positions, decision frameworks and boundary examples consulted on nearly every turn live in the wiki; detailed reference material, individual product specifications, historical decision records and niche external knowledge are served through semantic or lexical search.

The post cites Andrej Karpathy's LLM Wiki and Google's Open Knowledge Format, noting the shared insight that knowledge should be pre-extracted and explicitly structured rather than re-derived per query. For the reasoning side, Meta encodes expert methodologies as composable, imperative procedures it calls recipes, in contrast to the declarative knowledge files. The team says the patterns generalize to any enterprise domain with deep specialist knowledge, including finance, security and engineering.

Full write-up at Engineering at Meta.