Skip to main content
An attached causal graph gives your Agent two capabilities:
  1. Domain-grounded answers — the Agent can query the graph for relationships and reason with them.
  2. Automatic verification — every answer is checked against the graph before it goes out (see Verify agent answers).

​
How to attach

  1. Open the Agent block.
  2. Inspector → Data sources → Add causal graph.
  3. Pick the graph.
  4. Configure mode:
    • Reference — Agent can query the graph. No auto-verification.
    • Verify — auto-verification runs on every answer. Agent can query too.
    • Verify + rewrite — as above, plus the Agent retries on contradiction.

​
Tools the Agent gets

When you attach a graph, three tools are automatically added to the Agent:
  • graph_lookup(variable) — returns the variable’s type, range, description, and immediate edges.
  • graph_path(from, to) — finds the causal path (if any) between two variables.
  • graph_ask(question) — invokes an analysis session on the graph and returns the answer.
The Agent decides when to call these. In practice: graph_lookup for verifying its own claims, graph_path for “how does X affect Y” questions, graph_ask for anything requiring reasoning through multiple edges.

​
Multiple graphs

An Agent can have multiple graphs attached. Each is namespaced by its name so variable name collisions between graphs don’t cause confusion. Common pattern: one domain graph for the substance of what the Agent talks about, plus one operational graph for how your business runs (SLAs, escalation paths, ownership).

​
Access

Graphs respect workspace access controls. If a member can’t read a graph, they can’t attach it to an Agent they’re editing. Runtime access follows the same rules — an Agent running under a scope that lacks graph access falls back to unverified.

​
Editing vs. shipping

Attached graphs are pinned to a specific published version of the graph. Editing the graph in Draft doesn’t affect running Agents — you have to publish the graph, then publish the Agent’s Flow. Two publishes to ship a graph change. This is intentional: it means graph edits are reviewed at both the graph level (does this edge make sense?) and the Flow level (should this Agent care about this change?).

​
What to do next

Verify answers

Turn on auto-verification for high-stakes Agents.

Grounding

Use the same graph to cross-check retrieved chunks.