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
The first station is the message stack: your system message (the standing instructions), any history you preload, worked examples or an earlier conversation, and the newest user message, which usually arrives from an earlier block through a connection tag. The second station is the model: the brain that reads the whole stack. Sim supports the frontier providers, and because the block stays the same while the model changes, swapping brains never means rebuilding the step.
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 is the reasoning step of an AI workflow. Everything else in Sim's workflow builder is deterministic: triggers fire, blocks transform data, tables store it. The Agent block is where a language model reads the conversation you configured and decides what to write. Most AI agents you build in Sim are ordinary workflows with one or more Agent blocks doing the thinking.