Agentic Architecture: The Need For Knowledge Graphs & Structural Causal Models
One layer says what exists, and the other says what happens when you use it. Most systems only have the first.
A knowledge graph represents the system (entities and how they relate), and a structural causal model represents the mechanics (how the thing actually behaves when you act on it). Structure plus mechanism provides ‘what it is’ plus ‘what happens if’. Both are critical for going from descriptive agents to predictive, prescriptive, and diagnostic agentic capabilities.
Agents that just describe cannot be trusted to do work. You need all 4 capabilities of my information flywheel.
In this article, I will make the case for the framing, explain the differences between a knowledge graph and a structural causal model, and show you the systems that actually implement the framework I’m describing.
I am using research to dance around my NDAs so I can be more specific about the implementation than usual.
Structure Is Cheap. Mechanism Is Much Harder.
Let’s start with what each layer is, because the framework rests on them being different things. A knowledge graph is a map of entities and relationships. Nodes are things like customers, services, proteins, robot joints, and products. Edges are relationships like treats, depends-on, located-in, and purchased. This is the structural layer, and it’s enormously useful. It tells you what’s in the world and how the pieces connect.
ConceptNet and WordNet are complex knowledge graphs. Your CMDB is a lightweight knowledge graph. The org chart is a partial knowledge graph (taxonomy, really). None of these has enough substance to be an information model that supports reliable agents.
What a knowledge graph does not tell you is what happens when you reach in and change something. In the org chart, what causes someone to move from director to VP? It’s unclear, and applying causal discovery methods on historical data will describe what happened, but may not reveal the mechanics. That’s the part that matters most for decisions and actions, because each one is an intervention.
That’s the job of the second layer. A structural causal model (SCM) is the formalism Judea Pearl built to represent mechanisms. Strip away the notation, and an SCM is three ideas.