Aglet

Triage Agent Abstention Behavior

An agent can fail by guessing when evidence is missing or by refusing a request it could answer. Triage abstention around the evidence and action boundary. Include answerable, unanswerable, ambiguous, and conflicting-context cases before choosing a score for the boundary.

Establish what is happening

  1. Define answerability

    State what context, task information, and confidence support an answer. Identify when the agent should answer, qualify, ask a question, or decline. Keep unsupported certainty separate from a concise answer that omits optional detail.

  2. Map abstention cases

    Include complete evidence, missing evidence, conflicting sources, ambiguous intent, out-of-scope request, and safe partial answer cases. Record available context and expected user next action. Mark domain or policy limits that affect interpretation.

  3. Bound the score

    Specify whether the evaluation measures abstention choice, explanation, clarification, useful partial help, or calibration. Define over-abstention and under-abstention separately. Do not reward a generic refusal without checking whether it moves the task forward.

What to carry forward

The triage output is an abstention contract with answerability, evidence, response choices, user next action, and score boundaries. Stop when reviewers can distinguish prudent uncertainty from unnecessary refusal. Keep edge cases provisional when request intent is unclear in the current sample.

Technical background: Google DeepMind evaluation research.

Keep the decision with the work.

Use a Work Item in Aglet to record the problem, the evidence you have, and the next decision. Add an owner and priority, then keep updates in the discussion so the next person can follow the reasoning.

Create an account See the product workflow