Steps
Overview
The twelve step types, what each one does, and which ones can branch a run.
A workflow is built from twelve kinds of step. Each receives the previous step's result and passes its own along.
Core
| Step | What it does |
|---|---|
| Start | Receives the chat message and passes it on. A workflow needs one. |
| Agent | Calls a model from OpenAI, Anthropic or xAI. Can use tools, include earlier messages, and return text or JSON. |
| End | Returns the last output and stops. A workflow has one. |
| Note | A comment on the canvas. It is not executed. |
Tools
| Step | What it does |
|---|---|
| Tool call | Calls one tool on a built-in connector or an MCP server, with no model involved. Can require approval first. |
| File search | Searches the text files uploaded to the workflow. |
| Guardrails | Checks the message for personal data or jailbreak attempts, then follows Pass or Fail. |
Logic
| Step | What it does |
|---|---|
| If / else | Branches on a CEL expression. |
| While | Repeats the Loop path while a CEL expression is true, then follows Done. |
| Human approval | Pauses the run until a person approves or rejects. |
Data
| Step | What it does |
|---|---|
| Transform | Rewrites what the next step receives. |
| Set state | Stores a value that later steps can read. |
Note
A run stops after 20 step visits or when it reaches its token budget, whichever comes first. See Limits.