AnthropicClaude Certified Architect – Professional2026-104.3 Conduct A/B testing and iterative improvements
Prompt Versioning and Rollback
Completed by opening the official lab
Why it matters
When the prompt lives on the server rather than in code, a release is a change of version number. Rollback is only as fast as the pin that callers use.
What to practise
Create a support-ticket triage agent through the Claude Managed Agents beta API (client.beta.agents), then score version 1 on twenty labelled tickets through sessions pinned to {"type": "agent", "id": ..., "version": 1}. Ship a broad routing rule with agents.update, which creates version 2. Score again and watch billing accuracy drop, because billing tickets mention API usage too. Then roll back by pinning version 1. Note that a bare agent ID always follows the latest version and that agents.update has no approval step.
How it fits
Scoring both versions on the same fixed tickets is the attribution that 4.3.1 requires. Making the production pin the reviewed gate, with a canary slice before promotion, is 4.3.4’s staged release and restore. Compare the per-team result with 4.3.3 before you trust a small difference.