As AI products move from generating recommendations to taking consequential action, product leaders face a new operating question: who authorized the agent to act? This session gives product teams a practical playbook for defining bounded autonomy—who can delegate what, to which agent and tools, for which purpose, within what limits, and with which approval, audit, revocation, and recovery paths. Using an authority envelope, autonomy ladder, five-minute action reconstruction test, and paired adoption/control metrics, Sparsh will show how to write stronger PRDs, make permissions a product surface, design refusal and escalation behavior, and expand agent autonomy only when evidence supports it.