Workflow overview

The Flow canvas is where you build an agent. Drag blocks in, wire them together, and ship.

A Flow is one agent workflow, drawn on a canvas. You compose it out of a small set of blocks — each block is a unit of behavior that can be edited independently, wired to other blocks, and versioned.

The five block types

Block	What it does
Agent	Runs an LLM with a role, tools, memory, and data sources. The core reasoning unit.
Tool	An HTTP call or inline function the Agent can pick during a run.
Action	A side-effect that fires after the Agent decides — no LLM in the loop for arg selection.
Trigger	An entry point: chat message, webhook, schedule, or event.
Router	Branches control flow based on Agent output.

How a run flows

1. A Trigger fires (a user asks, a webhook comes in, a schedule kicks off).
2. The Flow enters the first Agent block. It reasons, pulls context from Memory and Knowledge, and may call Tools.
3. When the Agent decides an outcome, downstream Action blocks fire in order.
4. The full run is captured as a Trace you can inspect later.

Draft vs. published

Every Flow has two states:

• Draft — what you’re editing right now. Changes are live for you but not for real users.
• Published — the version production traffic hits. Frozen and immutable until you publish again.

Editing never risks production. Publishing is a deliberate step — see Publish a version.

Real-time collaboration

Multiple people can edit the same Flow at the same time. Cursors show who’s where, and edits merge without conflicts. You’ll see initials on blocks that a teammate is currently touching.

What to build first

Flow canvas

The workspace itself — panning, zooming, keyboard shortcuts.

Agent block

The most important block. Prompt, model, tools, limits.

Wiring

Connect blocks so data flows through them.

Publish

Ship the Draft as a new version.