Agent near-miss
Reviewed 2026-07-09 · v1.0
a working definition
An agent near-miss is an event where an AI agent's behavior diverged from the task it was actually given — an unusual request, an unexpected action, a proposal that did not fit the instruction — in a way that could have caused harm, whether or not any harm actually occurred.
what it looks like in practice
An agent mid-task asks for a standing password ‘to save time’, proposes emailing a file nobody mentioned, or insists on a step its operator does not recognize. Nothing has necessarily gone wrong yet — the human can still decline. The near-miss is the event itself: the moment the agent's behavior stopped matching the task, which is worth logging regardless of the outcome.
the machinery parallel
The dropped tool that hit no one. On a factory floor, a tool dropped from height that injures no one is still logged as a near-miss — not because harm occurred, but because the next drop might land differently. Agent near-misses are logged on the same logic: a caught or harmless deviation this time is data about what the system will do next time, under less attentive supervision.
how temja trains it
Temja's foundational course teaches the off-feeling as a signal worth acting on, and the shape of a useful report: sequence, root cause, contributing factors, and corrective action — deliberately without a name to blame, since a report that blames a person teaches the organization nothing repeatable. Temja's builder curriculum extends the same four fields into an incident-response runbook, where the corrective action is usually a change to the system's design rather than a reminder to a person.
questions
Does reporting a near-miss get the employee in trouble?
The opposite is the intended design: near-miss reporting is protective, not punitive. The person who flags an odd agent action is the one who kept a small event small. The failure mode training is built to avoid is silence, where an unreported near-miss leaves the same trap set for the next person.
What's the difference between a near-miss and an incident?
An incident caused actual harm — data left the building, money moved, a record was changed incorrectly. A near-miss is the same category of deviation, caught before or without that consequence. Both get the same four-field report; only the severity and the response differ.