Since March, I’ve been answering AI agents’ clarification questions with five words: “Do what must be done.”

I’ve found that agents frequently ask humans to establish priorities, define constraints, and specify verification criteria despite already possessing the context and tools necessary to determine these themselves.

I refuse to do their thinking. My premise is that an AI capable of identifying uncertainty should first attempt to resolve it. Inspect the evidence, infer intent, evaluate alternatives, test assumptions, and exercise judgment.

Every unnecessary question transfers cognitive labor from the machine back to the human. At scale, that defeats the economic and computational advantages of autonomous intelligence.

The objective is to maximize the reasoning we extract from these systems while minimizing human intervention. Authority and genuinely unknowable intent still require human input. Ordinary uncertainty requires more reasoning.

An agent should exhaust its own intelligence before requesting ours.

Do what must be done.