Connecting work with Arcs
Compose Machines into a visible workflow with decisions, operator review, delays, and outputs.
Start with the flow
An Arc is a directed workflow. A trigger introduces the input, Machine nodes perform tasks, and an output node defines the result. Decision nodes choose a branch; human review nodes pause for an operator; delay nodes wait before continuing.
Design the flow around explicit handoffs. A research step can produce a briefing, an analysis step can extract a decision field, and a review step can present that result to an operator. Each step should have a reason to exist.
References and decisions
Instructions and output templates can refer to upstream context. Give nodes clear labels or aliases so a later step can identify the result it needs. Decision conditions support comparisons, text containment, and checks for a field's presence.
When a decision depends on a number or boolean, prefer a structured field over trying to infer meaning from a paragraph. Review the true and false branches and make the resulting action explicit.
Validate before deploying
The builder validates the graph before execution. Resolve structural problems and unavailable Machine references, then use a test run to inspect the path through the graph. Deploy the Arc when it is ready to accept ordinary runs.
A run stores a graph snapshot, so its execution can be understood against the workflow that started it. Inspect node outcomes and run events when a branch fails or a review is waiting, rather than assuming the overall status explains the cause.
Review and failure paths
Choose what a Machine node should do when its task fails: stop the workflow, skip the step, or ask for review. A human review node has explicit approve and reject paths. Make both paths useful before attaching an automated trigger.
Arcs currently prevent a second active run of the same Arc. Account for that behavior when scheduling work or sending repeated webhook requests; a busy workflow can skip a new automated start.