Abstaining as an agent: when not acting is the correct output

What it is

Abstention is an output that says "no decision" and triggers a different path: a confirmation, a person, a stronger model, a retry with more context, or simply nothing. It applies to a classifier that picks a category, to a router that picks a handler and to an agent that decides whether to execute a step. Without it, every input is mapped to some action, including the inputs the system cannot read, and the error surfaces later as a wrong action rather than now as an open case.

Why it matters

A system that can only act is measured by the wrong-action rate, and that rate is dominated by the unclear cases. Giving the system a third output moves those cases out of the wrong-action count and into a queue that can be handled deliberately. Vendors of decision models describe the same structure: TypeSafe's confidence documentation recommends splitting confidence into three ranges (act, proceed with caution, do not act) and setting the boundaries by the stakes of each action, with a floor below which the system never acts on the answer. The principle does not depend on that model: any probability, score or self-assessment can gate an action, as long as the gate exists.

How to apply

• Add an explicit option for "none of these" or "cannot tell" to every closed decision, so that a poor fit is an answer rather than a forced choice.
• Put a floor on the confidence or probability below which the agent does not act, and separate thresholds per action that rise with the cost of a wrong action.
• Decide where abstained cases go before turning the automation on: a review queue, a confirmation prompt, a fallback rule. An abstention with nowhere to go is a silent drop.
• Count abstentions as their own outcome in monitoring; a rising abstention rate is an early signal of drift or of a new kind of input.
• For an agent's own steps, treat "I do not know enough to do this safely" as a legitimate completion state and report it as such, with what would resolve it.

Pitfalls

Thresholds set once and never revisited turn into either a wide-open gate or a queue nobody reads; both are measurable and both are common. Abstention that is punished (an agent judged only by completed tasks, a classifier judged only by coverage) is trained away in practice. And abstaining is not free: the person or system that receives the case pays for it, so the gate should be tuned against that cost, not only against the cost of a wrong action.