Add AI agents to a workflow
An agent is a step that uses an AI model to read, draft, classify or check. You give it instructions and the tools it may use. Everything else in the workflow stays predictable around it.
Choose the kind of AI step
| Step | Use it to |
|---|---|
| Agent | Do a defined job with tools, such as drafting a summary from three records. |
| Thinking agent | Work through a harder problem that needs a plan first. |
| Judge | Score another step’s output against a rubric before it goes further. |
Set up an agent
For each agent, set:
- Instructions: what the agent does and what a good result looks like.
- Model and effort: which AI model it uses, and how much reasoning it spends.
- Tools: the operations it may use on each connection. An agent can only use the tools you allow. For example, it can read deal records without being able to change them.
An agent can be shared between workflows. Fix or improve a shared agent once, and every workflow that uses it gets the change when it’s next published.
Keep a person in charge
Agents prepare work. People decide. Put a Person step after any agent whose output needs approval, and a Judge before it to catch weak drafts early. See Make a decision a workflow asks for.
Control the cost
Each run records what its AI steps cost. A step that reaches a spending limit stops with the outcome Capped, so a runaway step can’t spend without limit. AI spend from workflows shows in Flowstate Helm.