Agents

Most tools let you chat with an agent. Semantix lets you build one: drop the pieces on a canvas, wire them together, give them capabilities, and press Run.

The surface is called the Agent Composer, and the pieces in its palette are not a drawing of the engine — they are the engine. The list you drag from is served by the SDK itself. There is no separate "configuration" that gets translated into something real later. What you draw is what runs.

Opening it#

The composer is a full-screen mode of the IDE. In the left sidebar's tab row, click the Agent Composer button. The whole IDE body swaps out, under a slim bar reading Agent Composer on the left; to come back, use the IDE button on the right of that bar.

What you're looking at#

Components (left)	The palette. Six groups, served by the SDK.
The canvas (centre)	Where you draw. Dot grid, minimap, zoom controls.
Inspector (right)	Properties of whatever node is selected. Until you select one it reads "Select a node to edit its properties."
Flows (leftmost, optional)	Your saved flows. Opened from the header.

Above the canvas sits a strip with the flow's name, Run, Save and Export JSON.

Both side panels are drag-resizable and remember their width. Drag one narrow enough and it collapses to a strip you can click to reopen.

The palette#

Six groups, in the order they appear:

Agents — one entry, LLM Agent. The thing that thinks. → The agent

Pipelines — Linear, Parallel, Feedback Loop. How several agents relate to each other. → Pipelines

Triggers — Message Trigger. Where the flow's input comes from. Every flow needs one.

Outputs — Text Output, Events, Chat. Windows onto a run. They watch; they never execute. → Running a flow

Containers — Group and Flow. Dashed boxes that hold other nodes on the canvas. Purely visual: they never run, and dragging one moves everything inside it.

Tools — twenty capabilities you can hand to an agent: reading and writing files, searching your code, running an allowlisted command, and more. → Tools

⚑ The two gestures#

This is the one thing to learn before anything else, because the palette holds two opposite interactions and never says so in words.

Row	How you use it
Every component (agents, pipelines, triggers, outputs, containers)	Drag it onto the canvas
Every tool	Click it, then click an agent

A click on a component does nothing at all — not an error, not a hint, nothing. A drag on a tool does nothing either; tools deliberately carry no drag, because a tool never becomes a node.

The cues are there once you know to look. Tool names are in a monospace font with a round icon; components are in normal text with a square one, and show a grab cursor and a drag grip when you hover. Hovering a tool spells the gesture out: "Assign read_file to an agent".

⚑ Before you build anything#

Three facts that cost people work:

• There is no undo. No history, no Ctrl+Z. The only way back is to reload your last save.
• Delete is Backspace — not the Delete key. Nothing in the app says so.
• Deleting a container deletes everything inside it. Combined with the two above, that is the most expensive misclick here.

Save early. Saving and sharing explains where flows live and how they're named.

Where to go next#

Your first flow — trigger, agent, output, Run. Ten minutes, start to finish.

The agent — its settings, which models it can use, and what it does and doesn't remember.

Tools — the twenty capabilities, how you assign them, and what they're allowed to do.

Pipelines — running agents in sequence, in parallel, or in a loop.

Running a flow — what the canvas shows you while it runs, and how to read it.

Saving and sharing — where flows are stored, and how to hand one to someone else.