The canvas

Everything the workflow editor does with a mouse and a keyboard: getting around, wiring nodes, running, and reading what came back.

Workspace explains why you would draw a workflow instead of asking the assistant. This page is the editor itself — open any workflow from /user/workspace and you are looking at it.

What is on screen

A workflow is one canvas. Nodes sit on it, edges join them, and the whole graph is saved as a single thing you can run, share, or hand to a trigger.

Down the left is the node library, along the top the run bar, bottom-right the view dock, and the middle is yours.

Getting around

Zoom with the scroll wheel or a trackpad pinch. The percentage in the bottom-right dock is a button — click it, or press Cmd/Ctrl + 0, and you are back at 100%.

Pan by dragging any empty part of the canvas. Dragging a node moves the node; dragging the space between them moves the view.

Select one node by clicking it. Select many by holding Shift and dragging a box around them — or click the marquee tool in the dock if you would rather not hold a key. Whatever the box touches moves together.

Right-click a node for Duplicate and Delete. Right-click an edge to delete just that connection.

The view dock

Labels are the other rail button: drop a coloured note on the canvas to say what a cluster of nodes is for. They are annotation only — nothing runs them, nothing is billed for them.

Adding and wiring nodes

Open the library, search, and drag a node out — or click it and it lands in the middle of the view. The library is grouped in tabs (Text, Media, Control flow, Actions, Apps); the Apps tab only lists connectors you have actually connected, so you are never offered a node that can only fail.

Wire by dragging from one node's output circle to another's input circle. Ports are typed — a text output goes into a text input — and a node missing something it needs says so when you run it rather than failing silently halfway.

Three shortcuts to a starting graph:

• Start from a template — the gallery at /agent-templates drops a working graph straight onto a new canvas
• Describe it — the Agent button in the run bar takes a sentence ("webhook lead → agent → Slack") and draws the first draft for you
• Duplicate an existing workflow and cut it down

Subflows

A Subflow node runs another of your workflows as a single step: its input goes in, that workflow's result comes back. When a canvas passes ten or fifteen nodes, or when two workflows share a middle section, lift the shared part into its own workflow and point a Subflow node at it. One thing to fix instead of two, and a graph you can still read.

Running

Run saves the canvas and executes it. Nodes colour as the run moves: green for done, red for the one that failed, dimmed for anything a branch chose not to take. Dimmed is not failed — a branch that skips half the canvas is a branch doing its job.

A few rules the run bar enforces before anything happens:

• One run at a time per workflow. Two runs of the same graph would race on whatever they both write to
• Credits are checked up front, against the worst case of every node executing. You are told you cannot afford the run instead of discovering it halfway with half the emails already sent
• Your team's ceiling on concurrent runs applies across everyone, not per person

You do not have to save first. The canvas posts its current graph, so an edit can be tried before you commit to it.

Billing is per node that actually executed. A skipped branch costs nothing; a run that dies at node nine still pays for the eight that did real work.

When it fails

The run stops at the failing node and tells you which one and why — not the line number it stopped on. Every node's inputs, output, duration and error are kept, so a twelve-node graph that broke points at the break.

Run it again and it resumes from what did not finish rather than repeating the work in front of it. For a workflow that already sent email, that is not a cosmetic difference. Transient failures — a model call, an HTTP request — retry on their own; nodes with side effects never do, because a duplicate send is worse than a failure.

Previous runs

The clock icon in the run bar opens the history: when it ran, its status, how many of its nodes finished, how long it took, and what set it off if it was not you.

Click any past run and its node statuses are painted back onto the canvas, with whatever it produced opened alongside. Statuses alone tell you a run happened; this tells you what came out of it.

Triggers

A Trigger node is how a workflow runs without you. Set its kind:

• schedule — hourly, daily, weekly, or a cron expression
• webhook — an external HTTP request starts the run and its body becomes the input
• form — a public form submission (below)
• email — an arriving message

Everything downstream of the trigger runs on that event. Terminal outputs are filed in Artifacts when the run succeeds.

Share as a form

Form turns a workflow into a public page anyone can submit to, without an account and without seeing the graph.

Name the fields — each becomes a port on the trigger node — pick their types (text, textarea, email, url, number), mark what is required, then copy the link. Optionally the submitter is shown the result when the run finishes. The submission count sits in the same dialog.

Fields are typed here rather than inferred from the graph, because a graph is usually wired before anyone has decided what the form should ask.

Visibility

A workflow is private to you or shared with your team. Team members open it, run it, and edit it; a workflow published as a template is a copy, so nobody editing theirs can touch yours.

Shortcuts

Worth doing

Save before you run. A run is recorded against a saved workflow, so an unsaved canvas leaves you with a result and no history of it.

Name nodes for what they do, not what they are. Six nodes called LLM is a graph you will re-read every time you open it.

Split at ten to fifteen nodes. Past that, lift a section into a Subflow.

Test in pieces. Run, fix the node that went red, run again — the resume means you are not paying for the front of the graph each time.

Auto format after big edits, so the shape on screen still matches the shape of the logic.

Put a Slack or Email node on the failure path of anything running on a schedule. A workflow nobody is watching should say when it stops working.