Stop Telling Your Agents What to Do

The conversation

An agent that can choose only the next prescribed step has little agency to exercise. Yet replacing that workflow with an open-ended goal leaves a harder question unanswered: who decides what the agent is allowed to do?

Boundary-Driven Autonomy is Matt Vegas' framework for placing human judgment at the boundary-design layer. Specify the objective, context, permissions, prohibited actions, resource budgets, success criteria, and escalation conditions. Then give the agent room to choose its execution path. Permissions and spending limits need enforcement outside the model; a prompt cannot grant itself authority.

A hypothetical incident investigation makes the distinction concrete. The agent may gather evidence and compare fixes, but cannot change production, widen access, or exceed its budget. The design question is which decisions can be delegated and what evidence permits the next consequential action.

Prescribed workflows remain useful when the path itself is a requirement. If success cannot be checked, a boundary cannot be enforced, or a mistake cannot be contained, the right answer may be less autonomy.

The audience leaves with a way to separate instructions from controls, design autonomy zones, and place human review where judgment changes the risk or the decision.

Define the boundaries. Don't dictate the path.

What your audience takes away

• Separate the objective an agent must achieve from the execution path it may choose.
• Define autonomy zones through permissions, prohibited actions, budgets, validation, and escalation.
• Enforce authority and resource limits outside the model instead of relying on prompt compliance.
• Place human review before consequential decisions whose evidence or authority remains unresolved.
• Recognize when a fixed workflow, narrower tool access, or no agent is the better design.

The frameworks behind the talk