On the canvas, an Agent block looks like any other step: a run arrives at its input handle, and a result leaves on the other side. Inside, it's a conversation machine. This video follows the data in, station by station.
What you will learn
The message stack
Every run starts from a conversation: your system message, any history you preload, and the newest user message: filled in from the blocks before it.
A swappable brain
The Model field picks which provider and model read the stack. The block around it never changes, so you can hand the same conversation to a different brain without rebuilding the step.
Tokens in, tokens out
Everything on the stack is sent to the model and billed as input; the reply streams back and is billed as output. Tools attached to the block can be called mid-run.
The output bundle
Every run ends in a recorded bundle, the content, the model, the token counts, and any tool calls, that later blocks read through connection tags.
Here is the workflow the video rides through: the Qualify agent between its input and its result:
What's inside the block
Configure the system message with standing instructions, add any user and assistant history, and pass the newest user message through a connection tag. Choose the provider and model in the Model field. You can change the model without recreating the block.
What a run costs and returns
When the run reaches the model, every message on the stack ships as input tokens; the reply streams back as output tokens. If tools are attached, the model can call them mid-run. Everything the run did, the content it wrote, the model that ran, the token bill, the tool calls, lands in the block's outputs, where any later block can read it.
Where it fits
The Agent block runs a language model with the messages and tools you configure. A workflow can combine several Agent blocks with integrations and control-flow blocks. Other blocks, such as Router, can also use a model.