Definition: Human approval
Human approval is a workflow step — { "human": { "title": "..." } } — that pauses a run until a named person decides. The decision, the decider, and the record as shown are stored in the run, and the workflow resumes with the outcome as data.
How the pause works
Steps before the human step assemble the case — extraction, scoring, checks — and the step presents it under a plain title: “Approve invoice for payment”. The run waits at no compute cost. The approver can approve, reject, or request changes, which returns the run to a named earlier step with feedback attached and the attempt counted.
Details that make it a control
A human step can show a specific field from an earlier step’s output, declare which fields the approver may edit, set an expiry (expires_in_seconds), and notify people as it waits. On rejection, on_reject names where the run goes. These are declarations in stack.json, so “who can approve what, with what evidence, for how long” is reviewed like code — because it is code.
Frequently asked questions
A titled task with the context the workflow shows them — a field from an earlier step (the extracted total, the flagged clause) and the record behind it. A bare “approve?” button is a declaration failure, not a platform limit.
Yes. expires_in_seconds bounds the wait, and steps declare what happens on rejection or requested changes — return to a named step, notify, escalate. Absence handling is written down in the workflow instead of living in someone’s calendar.
A waiting run burns no compute. It resumes when the decision lands, which is why week-long approval chains are economically boring — the meter only moves when functions or workers execute.
Related capabilities, architectures, and comparisons
Ship your stack on the Software Lifecycle Cloud.
Closed preview on dedicated European infrastructure. Bring a repository or start from a blueprint.