AI Agent
Hands a task to an AI agent as a step in the flow: it reads a prompt, calls whatever tools it has been given, and returns a message and optionally a structured result. Reach for it when the work needs judgment or language—summarizing, drafting, classifying, reconciling two systems—rather than a fixed transformation.
The agent runs on the automation server itself, which is what lets one step combine tools from several environments in a single run—reading from one system and writing to another.
Fields
| Field | Description |
|---|---|
| Agent | The agent this step runs as: its model, its standing instructions, and its tools. The prompt below is the task on top of them. Leave it blank where the deployment permits, and the step runs on the deployment's own AI configuration. |
| Prompt | The natural-language task. Accepts variables, so it can work on the trigger's data and earlier results. |
| Payload Type | An output shape the agent fills in as structured data, alongside its message. Choose Content Type to take the fields of an existing content type, or Custom Shape to define the fields here. |
| Human in the Loop | Opt in to let the agent pause and ask a person before finishing. Choose the Providers the question is pushed to; Only Notify on Questions quiets the agent's running progress so only the moments it actually pauses are pushed out. Left unset, the agent must finish on its own. |
Overrides
| Field | Description |
|---|---|
| Model Override | Run this step on a different Provider Account and Model than its agent's, with an optional Reasoning Effort. Leave unset to inherit. |
| Ignore Agent Instructions | Run this step without the agent's standing instructions. The prompt still applies, and so does the framing this step type always imposes. |
| MCP Servers | Tool sources this step adds to its agent's. Add more than one to let the agent work across systems in a single run. A step can only add tools, never take the agent's away. |
What later steps see
| Value | Description |
|---|---|
message | The agent's text response. |
payload | The structured result, when a payload type is set. |
Which model it runs on
The model is resolved in three tiers, most specific first: the step's own Model Override, then the agent's own model, then your Default Model. The first tier that is filled in wins, and an empty tier falls through to the next.
A tier that is filled in but cannot be resolved fails the run—it does not quietly fall through. Naming an account that has since been deleted, or one whose credentials are incomplete, fails the step and says so, rather than sending your content and your costs somewhere you did not choose. With nothing configured at any tier, the run fails saying no model is configured for the step, its agent, or as your default.
You do not have to work the tiers out yourself. Where a step departs from its agent, a note at the top of the form says how—the model and which tier chose it, whether the agent's instructions are suppressed, and how many tools the step added. A step running its agent as configured shows no note, because there is nothing to report. When the configuration cannot be resolved at all, that note becomes a red warning carrying the message the run would fail with.
For agents, provider accounts, and how a model choice is made, see AI agents and models.
MCP servers
An MCP server is a source of tools an agent can call. Which ones a step may draw from is a property of the step: its agent contributes a set, and the step's MCP Servers field adds more on top. The servers themselves are set up once on the Connections page, and there are two types:
- Brightspot—the MCP endpoint of a Brightspot CMS. Takes a Name, the MCP Endpoint URL the CMS serves MCP at, and an Auth of OAuth or API Key, matching how an administrator configured that endpoint. With OAuth, clicking Connect discovers the endpoints and registers a client for you; with an API key, you paste the key and save.
- Custom—any remote MCP server, reached by its Remote MCP Server URL. A public server needs nothing else. For one behind OAuth, switch on Requires OAuth and click Connect: the endpoints are discovered from the URL, and a client is registered dynamically unless you supplied a Client ID and Client Secret for a pre-registered one.
A server's Name is how its tools are identified to the agent, so give it one you would recognize in a run log.
Human-in-the-loop
With Human in the Loop set, the agent can stop mid-task and ask a person a question—to confirm a decision, choose among options, or supply a missing detail. The step waits, the request is pushed to the providers you chose, and the agent resumes with the answer once someone responds. Every request is also answerable from the approvals inbox, whichever providers are configured—including none.
Gotchas
The node's Display tab Description, left blank, falls back to the opening of the prompt, so a card with a long prompt still reads as something. Set a description of your own when the prompt's first line is not what you want on the canvas.