If unsure, ask. Never guess. — AI Agent Pre-Execution Checklist

Some readers may raise concerns about this protocol. This document addresses them directly.

Concern 1: Isn’t having AI interpret a natural language Checklist still inference — the very thing this protocol tries to prevent?

This document does not reject AI inference.

The problem is not inference itself. The problem is executing on inference when the answer is unknown.

Reading a Checklist and determining what is known and what is not — that is permitted inference. Filling an unknown with a guess and executing — that is what this protocol stops.

Concern 2: Doesn’t requiring a Fixed Checklist contradict the flexibility of free-form JSON?

No. They operate at different layers.

The Fixed Checklist is the minimum structure for identifying unknowns. Free-form JSON is how known answers are recorded and processed.

There is no conflict.

Concern 3: Who generates the questions? If AI generates them, the same problem reappears.

AI does not generate the required questions.

The Checklist is built by the parties responsible for the Action.

At minimum, the Provider and the user define the questions. Where ethical, legal, or safety-critical items are involved, AI developers, organizations, or regulators may also contribute Checklist items.

This remains an open design question for the ecosystem.

AI developers who have sufficient confidence in their model may also choose to reduce the scope of the Checklist. The structure allows both expansion and contraction.

What does not change is this:

AI only determines which answers are already known and which are not. Unknown items are answered by the Provider in advance or by the user at runtime. Final approval always belongs to the user.

Concern 4: How is Human in the loop guaranteed?

It is not added as a separate mechanism. It is built into the structure.

C1 and C2 can only be answered by the user. AI cannot generate the user’s intent or execution meaning on the user’s behalf.

By following this protocol, human confirmation is structurally required before execution — not as an optional approval layer, but as a consequence of how the Checklist is designed.

Confirming intent means the human is already in the loop.