# Create a workflow

> Chain agents and tools into a repeatable, observable process with the visual workflow editor.

Source: https://cotera.co/docs/guides/workflows/create-a-workflow

---

Workflows turn a one-off chat into a process your org can rely on: defined steps, defined inputs, a run history you can audit. Build them on the canvas, or describe the process to Coco and edit what it drafts.

## The pieces

* **Nodes** are steps. A **tool node** performs one action directly (append a row, send a message, fetch a URL). An **agent node** brings judgment (read, decide, write). Control nodes (**filters, branches, delays**) route the flow.
* **Edges** carry data. Each node's output is available to the nodes after it, which is why structured outputs on agents matter: downstream steps address fields, not prose.
* **Triggers** decide when it runs: [on a schedule, from a webhook, or manually](https://cotera.co/docs/guides/workflows/triggers.md).

![](https://cotera.co/media/docs/deploy/deploy-workflow-canvas.png)

## Build one

1. Go to **Workflows**, create a workflow, and name the job it does (a workflow should do one job).
2. Add the trigger first: knowing the input shape (a webhook payload, a schedule with defaults) keeps every downstream step honest.
3. Add nodes for each step. Prefer tool nodes wherever the step is deterministic; use agent nodes only where the step genuinely needs to think. Most good workflows are mostly tool nodes.
4. Run it manually a few times from the editor and watch the run detail view: every node's input, output, and timing is recorded.
5. Turn on the real trigger when the manual runs look right.

## Iterating

Runs are the feedback loop. When a workflow misbehaves, open the run, find the node whose output surprised you, and fix that node: the run history makes regressions visible immediately.