Citadel Securities: Agents That Build a Model of the Company

How Citadel Securities Gives Agents a Model of the Company

Every company has a person who has survived 17 reorgs, knows the politics and somehow predicts how strategic moves will play out. Joao Wenzel of Citadel Securities opens his AI Engineer talk with that person: "That person is a company's world model, essentially." The problem is that they don't scale and nobody writes down what they know. At Citadel Securities, where more than two-thirds of employees write code and agents already help across the organization, leadership asked whether agents could help define and refine strategy, not just write code. That requires an institutional memory layer: who stopped talking to whom after a reorg, why a project was re-scoped, what people say and what they leave out.

Top-down sources lie politely: Decks, strategy docs and status reports are tidy but carry their authors' opinions, which produces what Wenzel calls watermelon projects: "They're green on the outside and red on the inside." Even extracting a status from a slide is hard. Parsing the PowerPoint XML raises the question of which hex code counts as green, and a vision model squints at a pasted screenshot trying to tell 31 from 81. The working answer combines both. Bottom-up sources, meeting notes, Slack, code changes and agent session traces, are noisy but closer to the truth.

A state estimator, borrowed from robotics: Wenzel's design treats the company model like a robot's belief about the world: confidence in a fact decays with time since the last supporting evidence and snaps back when a new deck, thread or meeting note arrives. Updates weigh how old the signal is, whether its source is still trustworthy and whether it agrees with other sensors. The loop is update, age, observe, grade and correct. Agents propose changes at every stage, while code and people decide how records are updated. Plain retrieval fails here: asked whether a project is on track, a RAG system happily reads a few Slack messages and slides and says green.

Three years of records and what they show: Each month the system combines last month's record with the new deck into a longitudinal record without rewriting earlier months, which makes later evals possible. It resolves projects that changed names by who was involved and continuity of goals, and it downgrades a green status to yellow when a comment mentions an outstanding dependency. Facts have a half-life: "on average, half the delivery dates change within two months, 90% of those change into a date that is after what was originally proposed". Letting one model read the whole corpus at once lost detail; a workflow of specialist agents reviewed by other agents more than doubled the number of insightful project updates.

Backtesting the company: Because records are dated, the team can snapshot the model as of February and compare its flags with what actually happened. Statuses that were green but flagged by the agent "were twice as likely to actually turn yellow or red." It doesn't predict exactly what goes wrong, but it tells leaders where to look. Human correction runs where people already work, in Slack, email or their apps, and the system distinguishes a corrected fact, which fixes one output, from a corrected rule, which fixes every output it covers. A quiet Slack channel is treated like a dead speedometer: missing signal, not a zero.

What it is used for: Leadership meetings start where the deck and the evidence disagree, managers get a weekly read on what stuck and what went quiet, and every other agent gets a shared context layer. Citadel Securities is extending the pattern to its client franchise. Open problems Wenzel names: matching one client across four systems with no canonical ID, and deciding when to apply an update automatically versus waiting for evidence or asking a stakeholder.

5 Takeaways From Citadel Securities on Institutional Memory for Agents

• Treat the company model as a state estimator - Every claim carries a date, a source and a confidence that decays over time
• Combine top-down and bottom-up evidence - Decks say what people report; Slack, PRs and agent traces show what is happening
• Never rewrite history - Feed-forward monthly records make backtests and evals possible
• Grade against what happens next - Green-but-flagged projects turned yellow or red twice as often
• Ask humans for rules, not just facts - A corrected rule fixes every output it governs, now and later

Why a Company World Model Matters for Enterprise AI Agents

Wenzel's warning is aimed at teams that think indexing all their documents gives agents context: "Don't assume that just like indexing everything and cataloging everything is enough". The information that matters lives between the lines and goes stale. Companies that start collecting dated, sourced records now will be able to test their agents against what actually happened, and that track record is what makes a strategy agent worth listening to.