Causal graph overview

A causal graph is a diagram of the variables in your domain and the causal edges between them. It looks like a network — nodes are things you can measure (rates, temperatures, prices), edges are the causal links between them. You use it for three things:

1. Explain agent reasoning — “the Agent recommended X because A causes B causes X.”
2. Verify agent answers — flag outputs that contradict the graph.
3. Analyze interventions — “if we changed X, what would happen to Y?”

​

When to build a causal graph

• Your Agent answers questions where being wrong has consequences — finance, medical, engineering.
• Your domain has well-understood cause-and-effect you can capture.
• You want Agents to cite mechanisms, not just correlations.

If your domain is fuzzy or the domain expertise is tacit, start with knowledge and memory. Add a causal graph only when you’re sure of the structure.

​

Graph shape

• Nodes are variables: continuous (blood pressure), discrete (department), or binary (fraud yes/no).
• Edges are causal: an arrow from A to B means “changes in A cause changes in B”.
• The graph is a DAG (directed acyclic). Cycles are rejected — they don’t make causal sense.

​

Building a graph

Two main paths: Most workspaces mix: seed with extraction, refine by hand.

​

Using a graph

​

Draft & publish

Graphs are versioned. You can edit in Draft (staged changes shown as pending), and publish when ready. Only published graphs are used in verification. Rollback is instant.

​

What causal graphs are not

• Not a replacement for knowledge. Keep your docs.
• Not a symbolic AI system. Your Agent still uses an LLM; the graph is a check, not the reasoner.
• Not a general knowledge graph. It’s specifically for causal relationships. If you want general entity relationships, use memory graphs.