You are releasing a new orchestrator prompt and tool schema for a production multi-agent concierge. Risk of quality regression is moderate; you need limited real traffic exposure with fast rollback. Which release methodology fits best?
Select an answer to reveal the explanation.
Short Explanation
You want a canary or progressive release: send a small slice of real traffic to the new orchestrator prompt and tool schema, monitor quality, error, and cost metrics, then expand or roll back fast. Big-bang cutover of 100% traffic with no rollback is too risky, and laptop-only deployment or unversioned live prompt edits fail production standards.
Full Explanation
Choose canary or progressive delivery for moderate-risk agent changes: route a small cohort of real traffic to the new multi-agent version, compare evaluation and operational KPIs such as quality, error, and cost against the baseline version, then promote or roll back. Keep artifacts versioned, including prompts, graphs, and tool schemas. The big-bang cutover of 100% traffic with no rollback plan is incorrect because an all-at-once change maximizes user impact if a regression occurs. Deploying only to an engineer's laptop as the production environment is incorrect because developer machines are not resilient multi-tenant production platforms. Changing production by editing live prompts manually with no versioning is incorrect because unversioned manual edits break auditability and safe rollback. Blue/green is also valid when you need an instant environment swap; canary emphasizes percentage-based exposure.