# Your Company Has No Model of Itself. Agents Can Build One - Joao Wenzel, Citadel Securities

## Executive summary

The video proposes building an 'institutional memory layer'—a world model for a company—using an agentic system that moves beyond simple Retrieval-Augmented Generation (RAG). This system approximates the knowledge of a single, omniscient employee by synthesizing evidence from disparate sources (e.g., decks, Slack, PRs, meeting notes). Drawing inspiration from robotics, the model functions as a state estimator, continuously updating its beliefs by incorporating new evidence, allowing facts to age, grading predictions against outcomes, and accepting human corrections. This approach helps identify critical project risks, such as 'watermelon projects' (green on the outside, red on the inside).

## Key takeaways

- The World Model as a State Estimator: The system must treat corporate knowledge not as a static database, but as a dynamic state estimator. This means tracking how long a fact has been observed (half-life) and how that signal is corroborated by other sources.
- The Core Agentic Loop: Effective knowledge capture requires a continuous loop: Update (incorporating new evidence), Age (applying recency bias to facts), Observe (using specialist agents instead of a single pass), Grade (back-testing predictions against actual outcomes), and Correct (allowing human feedback to steer the system).
- Bottom-Up vs. Top-Down Evidence: While top-down sources (decks, status reports) are clean but prone to human bias, bottom-up sources (Slack, PRs, agent traces) are noisy but contain high signal, requiring synthesis to build a complete picture.

## Technical details

- Data Synthesis and Evidence Sources: The system must integrate multiple data types: formal documents (decks, strategy docs), communication logs (Slack, meeting notes), code changes (PRs), and detailed user/agent traces. The synthesis of these disparate sources is key to building the 'institutional memory layer.'
- State Estimation Principles: The system must manage belief decay: the less frequently a fact is observed, the less trustworthy it becomes. This requires establishing a longitudinal, traceable audit layer for all claims.
- Specialized Agent Workflows: Instead of relying on a single large model pass (YOLO method), the system should use a series of specialist agents to break down collateral and look for specific details, significantly improving the quality of updates.
- Predictive Capabilities (Grading): By establishing historical snapshots (e.g., as of February), the agent can back-test its predictions, showing that flagged statuses (green but flagged) were twice as likely to slip or change color (yellow/red) than those left unflagged.

## Practical implications

- Automating the construction of a corporate 'world model' to guide leadership conversations by highlighting discrepancies between reported status (decks) and underlying evidence (Slack, PRs).
- Implementing a structured feedback loop (Correct stage) that incorporates human input to refine the model's understanding of rules, not just facts.
- Using the system to proactively identify high-risk projects that are superficially 'green' but contain underlying dependencies or issues.

## Topics

Agentic Systems, Institutional Memory, State Estimation, Knowledge Graph, Data Pipelines, AI Engineering, Citadel Securities, AIE CODE summit

Source: https://www.youtube.com/watch?v=jDwc0aXTAFg
