Pick two to five agent cards from the library, or paste your own, and write the objective. The composer returns a step plan: which agent runs which skill, in order, with the permission each hop still holds and the points where a human reviews before anything is written or spent.
The badge starts with what the whole job needs and narrows after every step, so a later agent cannot reach back for an earlier one's scope. The plan renders as a preview of the record the run would write, and opens the closest scenario in the sandbox.
This is a sample planner: deterministic, no model. Everything runs in your browser. We store nothing from your input except an anonymised event: which tool ran and when.