Your first flow
Trigger, agent, output — the smallest flow that does something, built one step at a time.
The smallest useful flow is three nodes: something to start it, something to think, and somewhere to put the answer. This page builds exactly that.
Before you begin, open a project. A flow whose agents carry no tools will run without one, but everything you'll want to do next needs a workspace, and the run refuses without it.
1. The trigger#
Drag Message Trigger from the palette onto the canvas.
Select it, and in the Inspector on the right find the Message field. Type what you want the agent to do — "List the top-level folders in this project and say what each one is for" is a fine first test.
Important
The Message Trigger is the input box. There is no prompt bar anywhere in the composer — no field above the canvas, no dialog when you press Run. What you type into the trigger's Message field is what the flow receives. The trigger itself never executes; it exists to hold that text.
2. The agent#
Drag LLM Agent onto the canvas, to the right of the trigger.
Select it and set two things in the Inspector:
- Model — the picker lists the models you've configured in Semantix and starts on the first of them. If you have none, add one in Settings first.
- System Prompt — how the agent should behave. Leave it blank for a first run if you like.
Everything else has a working default. The agent covers the rest of the fields.
3. Wire them#
Drag from the small handle on the right edge of the trigger to the handle on the left edge of the agent.
Note
The arrow decides the order. Nodes at the top level of a flow run in the order your wires describe. Anything you leave unwired falls back to the order it was created in — not its position on screen, which is rarely what you meant once you have more than one agent. Wire what should happen in sequence.
4. Somewhere to see it#
Drag Text Output onto the canvas to the right of the agent, and wire the agent to it.
The box reads "Run the flow to see output…" until something happens. It's resizable — select it and drag a corner.
If you'd rather watch the machinery than the answer, use Events instead; if you'd rather read it as a conversation, use Chat. You can wire all three to the same agent. See Running a flow.
5. Run#
Press Run in the strip above the canvas.
The button turns into a red Stop. The agent's border lights up and a small pill appears at its
top-right corner, counting through the run: leg 1 while it calls the model, then a live character
count while text streams, then ✓ 1.4s when it finishes. Your Text Output fills in as the answer
arrives.
If something is missing, the strip tells you before anything runs:
- "Add a Message Trigger to start the flow."
- "Message Trigger has no message — type one in the Inspector."
Both stop the run before it starts. Errors during a run appear in the same place, truncated, with the full text on hover.
6. Save it#
Type a name in the box to the left of the Run button, then press Save. The button flashes Saved ✓.
Important
The flow's filename comes from its name. Two flows left at the default name overwrite each other, and renaming a flow saves a second copy rather than renaming the first. Name it before you save it. Saving and sharing has the details.
What to add next#
Give it hands. Right now the agent can only talk. Click a tool in the palette — list_dir is a
gentle first one — then click the agent. A badge appears on the node.
→ Tools
Add a second agent. Wire the first agent's output into a second one and the answer flows onward. For anything more structured than a straight line, use a pipeline. → Pipelines