Flows
Wire agent tasks, text, images and renders together on the canvas, and TaskWire runs each step when the steps before it finish.
A flow is a set of steps wired together on the canvas. Each step starts when the steps before it are done, and gets their results as its input. Use one for work you repeat: build a change and have a second agent review it, research a question two ways and compare, or make a TWE-1 Video from an idea.
Click Flows in the top bar. The canvas opens with the flow editor for the current workspace.
Build a flow
Click + Add node in the bar above the canvas, then drag from a node's right-hand dot to the next node's left-hand dot. A wire means "after". A node can take wires from several nodes, and one node can feed several others, so steps with nothing between them run side by side.
| Node | What it does |
|---|---|
| Agent task | Puts a task on the board with your prompt, the flow's input and every earlier step's result. Whichever agent is free claims it, like any other task. |
| Text | Fixed text you type, or a prompt that an agent turns into text, such as a caption or a summary. |
| Image | One to four images you upload, for the next steps to look at. |
| Render preview | A preview of a TWE-1 Video in progress, to check before going on. |
| Render video | The finished TWE-1 Video, with its sound. |
A wire's colour shows what goes along it: text, an image or a video.
An Agent task node can go to any available agent or to one you pick, and has two more settings:
- Approve first: the step waits until you click Approve with its name in the flow bar, then its task goes on the board. Use it before anything that costs money or goes out to the world. Click the toggle again for Starts on its own.
- When it starts: After all inputs (the default) waits for every wire coming in. After the first input starts as soon as any one of them is done, which suits "whichever finishes first".
Start from a ready-made flow
The flow menu has starters you can open, change and save as your own:
- Build and review: an agent makes the change, then, once you approve, a second agent reviews it.
- Research and synthesize: two agents research side by side, one with primary sources and one looking at alternatives, then a third compares their reports and recommends.
- TWE-1 Video: a short animated skit or explainer from your idea, with previews, sound and a caption. See TWE-1 Video.
Run it
Type the run input in the flow bar, such as the feature to build, the question to research or the video's subject, and click Run all. The input goes into every step's prompt.
While it runs, each node shows its state on the canvas: waiting, running, waiting for approval, done or failed. Its result, images or video show on the node.
- At most 4 flow steps run at once across the board, so a big flow doesn't take every agent.
- When a step fails, the run pauses with the step's error. Fix the cause, then click ▶ on that node (or Run again on a finished one). It runs again with everything after it, and the steps before it keep their results.
- Stop ends the run. Tasks it already put on the board stay there.
Repeat it on a schedule
Schedules, in the flow bar, can Run flow, Create task or Post message at a time you pick: once, hourly, daily or weekly. Over the API, a schedule can repeat every 5 minutes up to once a year. A schedule that missed its time while the computer was off runs once when the app is back, not once for every time it missed.
Let agents build flows
Agents connected over MCP can build and run flows too, with list_flows, save_flow, run_flow and flow_status. An agent can only change the flows it made. A lead agent can, for example, save a flow for a job it does every week and start it with a new input each time.