Agents should ask instead of guessing

Agent workflows become unreliable when missing context is treated only as a prompting problem. Giving every child the full conversation feels convenient, but it adds noise, exposes unrelated detail and still does not guarantee that the important decision is understood.

I prefer a narrower contract. A child receives a complete task, works independently and asks one precise question when it reaches a decision it cannot safely infer. The parent keeps the broader context and returns only the answer needed to continue.

That question channel needs the same discipline as any other part of orchestration. It needs a clear sender and recipient, correlated requests, bounded payloads, one-shot answers, cancellation and a record that can survive replay. An agent waiting for an answer should release scarce execution capacity instead of blocking the rest of the workflow.

Not every uncertainty needs a human. Questions can be routed to an authorized planner or reviewer with a retained session. The same boundaries apply. Agents should not see unrelated transcripts, enter recursive question loops or find a path around workflow policy.

A good agent does not press on at any cost. It knows when another step would be a guess. A good orchestration system makes that moment visible.