Run from a workflow
A single agent answers a question. A workflow chains agents and tools into a process: fetch the new signups, enrich each one, score them, post the hot ones to Slack. If your agent is one step of something bigger, put it in a workflow.
Build it
- Go to Workflows and create a new workflow, or describe the process to Coco and let it draft the workflow for you.
- Add your agent as a node. Its structured output becomes the input available to downstream nodes. This is why defining structured output matters.
- Connect the steps: tool nodes for direct actions (no LLM needed), agent nodes where judgment is required, filters and branches for control flow.
- Add a trigger: a schedule, a webhook, or run it manually while you iterate.

Agent step or tool step?
- Use a tool node when the action is deterministic: append a row, send a message, fetch a URL. It's faster and costs nothing in model usage.
- Use an agent node when the step needs to read, judge, or write.
A good workflow is mostly tool nodes, with agents only where thinking happens.
Watch it run
Every run has a detail view showing each node's input, output, and timing. When a run does something surprising, the trail is all there.