Architecting an AI Agent for Expert Knowledge Capture and Reasoning
This article from Meta Engineering details the architectural design of an AI agent dubbed an "Organizational Second Brain." The system focuses on capturing, structuring, and leveraging deep specialist knowledge within an enterprise, separating knowledge from reasoning. It features a self-improvement loop that compiles expert feedback into verified updates without model retraining, making it highly relevant for understanding how to build scalable and auditable AI systems that integrate human expertise.
Read original on Meta EngineeringLarge organizations frequently struggle to capture and scale expert knowledge, which often resides implicitly within subject matter experts (SMEs). This leads to inefficiencies, inconsistencies, and risks, especially in high-stakes domains like compliance or security. Meta's "Organizational Second Brain" addresses this by codifying institutional intelligence into an AI agent designed to make deep specialist knowledge readily available and persistently learn from expert feedback.
Core Architectural Principles
- Separation of Concerns: The architecture distinctly separates what the agent knows (knowledge system) from how it reasons (reasoning layer). This modularity allows independent updates and clear failure attribution.
- Structured Knowledge Architecture: Implicit expert knowledge is made explicit and structured into machine-readable files (positions, taxonomies, routing indexes, gateways). This contrasts with approaches that re-derive reasoning from raw documents at inference time.
- Self-Improvement Loop: Expert feedback is systematically compiled into verified, regression-tested updates to the knowledge system and reasoning layer without requiring LLM retraining, ensuring continuous learning and institutional memory compounding.
Four-Layer System Design
- Knowledge System: A structured, auditable repository of organizational stances, taxonomies, and routing information. Files declare dependencies (depends_on) and consumers (referenced_by), forming a bidirectional dependency graph for traceability and automated updates.
- Reasoning Layer (Recipes): Composable, imperative procedures that define multi-step analytical workflows. Recipes reference knowledge files but contain no domain facts, ensuring clear separation. This enables progressive disclosure, loading only relevant instructions and knowledge at each step, significantly cutting context window usage.
- Evaluation Framework: Gates every proposed change, ensuring quality and consistency through testing.
- Improvement Loop: Integrates expert feedback (corrections, escalations) into permanent updates to knowledge files and reasoning recipes, creating a continuous learning cycle.
Knowledge Partitioning Strategy
A key design decision is how to partition knowledge. High-density, frequently referenced, and evolving knowledge (e.g., organizational positions, decision frameworks) is curated in the structured wiki. Sparse, situationally relevant information (e.g., detailed reference material, historical records) is served via semantic or lexical search (RAG) to prevent system bloat and dilute attention, ensuring core reasoning remains grounded in refined, current knowledge.
Human-in-the-Loop Mechanisms
The system incorporates checkpoints (intermediate reasoning review by experts) and escalations (agent hands off ambiguous cases to experts). These mechanisms ensure quality control, provide critical training signals for the self-improvement loop, and build expert trust by making the agent's reasoning transparent and allowing human overrides.