Agent systems often treat human questions as interruptions.

The agent gets stuck, asks in chat, waits for an answer, and then continues.

That works in a conversation. It breaks in a run.

A human question should be control flow. It should have a shape, a scope, and a continuation.

Questions Need Shape

Some questions are too vague to unblock work.

bad: What should I do?
better: Should the public API keep the old name?

A useful question records the decision being requested.

question
  scope: public API
  blocked lane: rename admission
  options:
    keep old name
    rename and add compatibility alias
  default: add compatibility alias
  expires: before merge admission

The answer can then route back into the system.

Answers Have Dependents

An answer is not just text. It can unblock several lanes.

If the user says a public API must remain compatible, that answer can affect implementation, tests, documentation, and review.

The run should know which work depended on that answer. If the answer changes later, the dependent work becomes questionable.

Do Not Ask Too Late

Agents should not wait until the end to discover that they needed a human decision.

Questions belong near the uncertainty.

before writing code:
  should this behavior change be public?
 
before merging:
  should this risk be accepted?
 
before release:
  should this compatibility break ship?

The earlier the question is attached to the right region, the smaller the wasted work.

Questions Are Not Failures

A good question can be progress.

It narrows the system. It identifies the missing decision. It prevents agents from guessing across a boundary they do not own.

The failure is not asking. The failure is asking in a way that cannot be routed back into the work.

The Mental Model

A human question is a node in the run.

It has parents: the evidence and uncertainty that created it.

It has children: the work that uses the answer.

That makes human input part of the system, not an aside from the system.