Agent
Calls a model from OpenAI, Anthropic or xAI. Set its instructions, tools and output format.
The Agent step sends the incoming message to a model, lets the model call the tools you select, and passes its answer to the next step.
Model
Choose a provider and a model. The step is configured the same way for all three providers, so you can switch models without changing the rest of the workflow.
| Provider | Key |
|---|---|
| OpenAI | OPENAI_API_KEY |
| Anthropic | ANTHROPIC_API_KEY |
| xAI | XAI_API_KEY |
Each Agent step has its own model setting, so one workflow can use a small model to classify a message and a larger one to write the reply.
Instructions
The instructions are the system prompt for this step. State the output you want and what the model should do when it cannot produce it.
Tools
An agent can use tools from a connector or an MCP server. Connect the service from the step's settings, then pick which of its tools the agent may call. The model decides when to call them.
If you know exactly which tool should be called and with what, use a Tool call step instead. It calls one tool directly, costs no tokens, and behaves the same way every time.
Conversation history
By default an agent sees only what the previous step handed it. Turn on Include earlier messages to give it the whole conversation. Steps that reply to the user usually need this.
Output
An agent returns text, or JSON when the next step needs structured data, for example an If / else that branches on a field.
Calling a tool without a model
Choose No model, call the tool directly to skip the model entirely. The step then behaves like a Tool call.