When does a workflow start running on its own?

When you say so. Nothing promotes a workflow to autopilot on a counter or a score. The platform shows you what it has actually seen, how many clean runs, which branches have occurred, how complete the SOP looks, and you make the call.

Coverage is the honest gate, not volume. A two-branch job can be understood after two runs. A workflow whose awkward cases turn up once a quarter is not ready after twenty, and no threshold we could pick would know the difference.

When you do pull the trigger, codification is a walk-through rather than a silent compile. The platform goes through the procedure with you against real example data, asks about the places where the examples were ambiguous, and lets you pin the steps that must stay agentic or must always stop for a person. Seeing a real run go past is what tends to surface the rules nobody ever wrote down.

Everything the walk-through surfaces goes back into the SOP, not only into the code. The workflow ends up better documented as well as automated.