Known destination. Real boundaries.

If the destination and boundaries are known, the system should be able to drive. That’s the whole thesis. Everything else is what “known” costs.

Autonomy doesn’t mean removing structure — it means having enough context, authority, and boundaries that you don’t need to ask what to do at every step. Deci and Ryan’s self-determination theory backs this up for people; organizational research calls it “discretion within a job,” not freedom from oversight. Same logic applies to agents.

The word doing the heavy lifting in “if known” is known. A destination described in a sentence isn’t known — it’s known when it’s checkable. A boundary isn’t real until it’s documented and enforceable, not implied. Simon’s bounded rationality is the reminder here: you can’t skip the specification work and still expect good decisions.

Permission isn’t the same as capacity. Aghion and Tirole’s distinction between the right to decide and the real ability to decide matters more for agents than for people — granting an agent formal authority without data, credentials, or tools doesn’t produce autonomy, it produces a bottleneck with a new name.

Last piece: escalation has to be a designed path, not a personality trait. Edmondson’s psychological safety research says this needs structural support, not a culture memo — agents need to be able to surface uncertainty without being penalized for it, especially before irreversible actions.

Known destination. Real boundaries. Actual authority. Built-in escalation. That’s the whole system.

Full argument, citations, and the human parallel: Destination and Boundaries →

Leer en español — La versión corta en español.