The conversation around AI agents has been dominated by what they can do. Far less attention gets paid to the guardrails that should govern what they are allowed to do. In the latest episode, Sahil Agarwal gets to the heart of this imbalance with the DPACT framework, a name that stands for Delegation, Policy, Auditability, Context, and Time. It is a deliberate move away from the brittle, token-based access that has defined our current security thinking. For anyone building on top of large language models, this is not an abstract concern. It is the difference between an agent that fetches the right file and one that quietly deletes a production database because its permission token was too broad.
Agarwal's central insight is that we have been asking the wrong question. The industry has been obsessed with *who* the agent is, when the real question is *what* the agent is allowed to do in a specific moment. That is why the framework leans so heavily on bounded delegation and explicit context. A static API key cannot capture the nuance of a request that involves sensitive customer data or a financial transaction. By introducing time limits and audit trails, DPACT forces a level of discipline that feels foreign to the rapid, iterative nature of AI development. We would tell any engineer listening to stop thinking of this as a compliance checkbox. It is a design philosophy that makes agents predictable enough to trust with real work, not just demos.
The practical takeaway here is that security is becoming a product feature, not a backend concern. Teams that adopt this mindset will find themselves moving faster, not slower, because they will be able to delegate with confidence. Imagine a support agent that can issue a refund only within a specific dollar range, for a specific reason, and only during business hours. That is the kind of granular control DPACT enables. It turns the fear of rogue AI into a manageable risk profile. For the reader who asks, "Where do I start?" the answer is not to buy a new tool. It is to map your current workflows against the five pillars and identify where your delegation is unbounded. That single exercise will reveal more vulnerabilities than any red team.
What remains open is whether the industry will coalesce around a shared standard or fragment into proprietary silos. Agarwal has given us a strong foundation, but adoption will require vendors to expose these controls in a way that developers actually enjoy using. The detail to watch is how quickly identity providers and agent frameworks start baking DPACT-like semantics into their default settings. If they do, we will look back at the era of blanket tokens as a primitive time. If they do not, we will continue to see headline-grabbing breaches that were entirely preventable. The future belongs to agents that earn their authority one bounded action at a time. That is the standard we should hold every new release to.
