Skip to content

AI Agents & MCPs & AI Workflow Automation • (~400 MCP servers for AI agents) • AI Automation / AI Agent with MCPs • AI Workflows & AI Agents • MCPs for AI Agents

2.5 KB
Raw Blame
---
title: "Agents in Flows"
sidebarTitle: "Agents in Flows"
description: "Use an agent as a step, and act on what it returns"
icon: "sitemap"
---

Drop an agent into a [flow](/flows/building-flows) as a step. The flow handles what's certain, the agent handles the rest.

## What people use it for

<CardGroup cols={3}>
  <Card title="Research" icon="magnifying-glass" color="#8143E3">
    Looks up a lead and returns a brief with a fit score.
  </Card>
  <Card title="Triage" icon="inbox" color="#F59E0B">
    Reads an email, classifies it, sets the urgency.
  </Card>
  <Card title="Action" icon="bolt" color="#16A34A">
    Files the invoice, updates the system, posts the summary.
  </Card>
</CardGroup>

<Tip>
If you can write the rule in one sentence without saying "depends", it's a condition, not an agent.
</Tip>

## Run one of your saved agents

Add a **Run Agent** step from the **AI** piece, then pick the agent in **Agent**. Its instructions, tools and model come from the agent, so leave **AI Model** empty. An agent marked **Not published yet** is listed too, but a flow can only run it once it's published.

<Steps>
  <Step title="The step runs the published agent">
    Edits you make to the agent reach your flows once you publish them from the agent's Configure panel. Until then, flows keep running the version that was live.
  </Step>
  <Step title="Each run shows up on the agent">
    The agent's **Runs** tab lists the runs flows started, with the flow that triggered each one. A run refused before it started isn't listed.
  </Step>
  <Step title="A published flow keeps the agent in place">
    While a published flow runs the agent, it can't be deleted or moved to another project. The dialog links the flows that use it (members who can't read flows see a count instead). Take the agent out of each flow and publish the flow, then try again.
  </Step>
</Steps>

## Get a result you can build on

Prose is hard to branch on, because the phrasing drifts every run.

> "This looks fairly urgent, someone should pick it up today."

Ask for **structured output** instead and you get fields:

| Field | Type | Example |
|---|---|---|
| `category` | Text | `billing` |
| `urgency` | Number | `4` |
| `needs_human` | Boolean | `true` |

Later steps pick these up from the [data selector](/flows/passing-data), the same as any other step's output. Route on `needs_human`, sort by `urgency`, drop `category` into the Slack message.

<Tip>
Ask for the fewest fields the next step needs. Every extra field is one more thing to get wrong.
</Tip>