As agents move from executing predefined tasks to making decisions and taking action, traditional access controls are no longer enough. Without clear context around data quality, provenance, permissions and refusal conditions, agents can act on the wrong information or take the right action in the wrong context. This talk explores how datasets can become active, governed control surfaces that communicate machine-readable rules directly to agents. Attendees will learn how to design data boundaries that give agents greater freedom to operate while keeping their decisions traceable, policy-aware and within defined limits. Key takeaways: → Why production agents often fail at the data boundary → How to expose semantics, provenance and quality signals agents can interpret → How permitted operations and refusal conditions constrain unsafe actions → How to increase agent autonomy without sacrificing control or accountability