Giving AI agents permissions and boundaries
By Robert OrozcoSeptember 19, 20265 min read
Agents are software that acts. Govern them the way you would govern a capable new employee with system access, not the way you govern a document.
Much AI governance was written for a world where AI produced text. Agents break that assumption. An agent does not just answer. It takes actions: it calls tools, moves data, triggers workflows, and sometimes spends money or changes records. Governing an actor requires different questions than governing an output.
The most useful mental model is a capable new hire with broad system access and no institutional memory. You would not hand that person every credential on day one and hope for the best. You would define what they can touch, what requires sign-off, and how you would notice if something went wrong. Agents deserve the same treatment.
What an agent boundary actually specifies
- Permissions: which systems, data, and actions the agent may reach, and just as important, which it may not.
- Escalation: which actions require a human decision before they happen, not a review afterward.
- Data boundaries: what the agent may read, retain, and pass to a third-party model or vendor.
- Accountability: whose responsibility an agent action is, recorded so it can be traced.
- Observability: how anyone would detect that the agent behaved outside its intended range.
None of this has to be heavy. The weight of the control should match the level of risk. An agent that drafts internal summaries needs light guardrails; an agent that can issue refunds or touch regulated data needs real ones. The failure mode to avoid is uniform governance: either so strict that nothing ships, or so loose that everything is permitted.
Usable beats documented
A governance framework that lives in a document no one opens is not governance. The test is whether the people building and using agents can follow the rules without a lawyer in the room. That means writing controls as defaults in the system where possible, teaching the people affected, and revisiting the boundaries as the agents’ capabilities change.
Done well, permissions and boundaries are not a brake on adoption. They are what makes confident adoption possible. Teams move faster when they know exactly what an agent is allowed to do and trust that the edges are real.
This piece relates to AI Governance.