Clarification-at-edge / Ask-before-act

Definition

Insert a clarification step at agent-agent handoff edges, or before uncertain actions, to reduce error propagation.

Category: Decision

When to use

Vague requirements, agent handoffs, long tasks, multi-constraint tasks, ambiguous user intent.

When not to use

When asking at every step is intolerable, or when the clarification cannot change the execution outcome.

How to implement

1. Score every cross-agent message for ambiguity / risk.
2. Above threshold, ask the source agent or user — don't guess.
3. Clarification questions should be few, specific, and actionable.
4. Record whether clarification actually reduced failure, and tune the threshold.

Minimal pseudocode

const score = ambiguityScorer.score(edgeMessage);
if (score > threshold) {
  const clarification = await askClarification(edgeMessage);
  edgeMessage = merge(edgeMessage, clarification);
}
return targetAgent.run(edgeMessage);

Recommended trace events

• clarification.triggered
• clarification.question.asked
• clarification.answer.received
• clarification.skipped

Common failure modes

• Over-interrupting the user.
• Questions too abstract to answer.
• Clarification results don't make it back into state.

Implementation checklist

• Input/output schemas defined.
• Each agent's permission boundary defined.
• Every agent call carries a run id / trace id.
• Failure, timeout, cancel, and retry strategies defined.
• Context passed is the minimum required, not the full history.
• High-risk actions are gated by approval or a verifier.

References

• AgentAsk