Two ways agents meet Actions
Agents and Actions connect in two directions:
This page focuses on Actions calling agents. For agents triggering Actions, see Focus & capabilities.
Use Agent: rails calling a driver
The Use Agent node invokes a configured agent without opening a conversational thread. You supply inputs — context, variables, prompts — and read the model’s output from the node for downstream steps (email, Slack, database updates, etc.). This is the right pattern when:- The step has a clear input and output shape
- You want agent judgment inside a guaranteed pipeline
- Downstream nodes need the agent’s text as data, not a chat UX
Full node wiring — variable slots, canvas connections, and a worked example — lives in Agent node. Read that guide when you build your first pipeline.
Chat vs. Use Agent node
Same agent configuration — different harness. Test conversational behavior in Chat or the Lab; test pipeline behavior in Action builder Test.
Variables in pipelines
When the agent’s system prompt declares variables, the Use Agent node exposes matching input slots on the canvas. Wire upstream node outputs into those slots — dates, names, counts, prior step text. Variable names must match what you defined in the Builder. Mismatched names produce empty placeholders and weak output.Where to put rails vs. judgment
From What agents are — and aren’t: use the Action graph for lanes and the agent for judgment.
The agent decides wording and classification; Actions guarantee delivery and enforce caps.
Diagram placeholder
Simple Action graph: Trigger → Use Agent → branch on output → Email node / Escalation node. User-facing node labels only.
Active version matters
The Use Agent node uses the agent’s active version at run time. Publish prompt changes via Save and confirm the right version is active in Version History before production Actions depend on them.Where to go next
Agent node
Step-by-step pipeline wiring with screenshots.
Agents in Tracks
Agents as overseers and task workers in workflows.