An agent must be able to say it does not know
Stopping and asking should be a recorded success, not a failure the system is tuned to eliminate.
“We hold that agents must be built and measured so that refusing to proceed on insufficient information is a valid outcome, recorded as such, never penalised as failure, and never trained away in pursuit of a completion rate.”
Any agent can be made to finish every task. You lower the bar for what counts as finishing. The output becomes confident, plausible and wrong, and the person who relied on it finds out later than they would have if the agent had said nothing.
Completion rate is the metric that creates this. Optimise it and you are explicitly rewarding a guess over a question, because a guess closes the ticket and a question does not. The fix is not to ask models to be humbler; it is to stop paying them for the wrong thing.
A stop is not a non-answer. 'I need the schema for this table before I can write the migration' is a result: specific, actionable, and far cheaper than the alternative. Systems should record it as the useful output it is.
Signing this means
- Count a justified stop as a completed run, not a failed one.
- Require a stopped run to say what it would have needed to continue.
- Do not rank or reward agents on completion rate alone.
Add your name
Signatures are counted per position, not per person. You can hold this one and not the next.
