The canvas

Each node is an agent running one of its modes. Drag from a node’s handle to another to connect them — that edge is the order of execution. Nodes carry their own mode and model pickers.

Use Load… to open an existing pipeline, or + New to start fresh. Save run def persists it.

Parallel stages

Stages that don’t depend on each other can fan out from one upstream stage and run at the same time, then rejoin at a later stage that waits for all of them. This is the difference between a critical path of sum of every stage and the slowest branch.

The classic shape is a diamond: one planning stage fans out to several independent stages (fetch an image, gather links, write the body), which all rejoin at a single publish stage.

A fan-in stage only runs once every upstream branch reports completed. A branch that soft-fails should still report completed with empty output, or it will skip everything downstream.

Scheduling

Scheduling lives on the pipeline, not the agent — a recurring job is just a pipeline with a schedule. That is why there’s no separate loop builder: a loop is a single-stage scheduled run.

Workspace scope

A pipeline carries a workspace_id. Pick a workspace to scope it there, or leave it global to run across all of them.