Do it with guardrails, and only after pre-production testing has cleared a basic quality bar. Cindy Sridharan's framing splits the work into pre-production, deployment, release and post-release stages. Feature flags separate deployment from release and act as a kill switch; canary releases expose a small share of traffic first; observability (metrics, traces, logs) tells you what the change is doing; and synthetic monitoring runs scripted checks against live endpoints. Set measurable thresholds and rollback triggers in advance, limit the blast radius, and use dedicated test accounts so live user data is not affected. Google's testing team makes the complementary point that monitoring is not useful if the service does not even build and pass its tests.

Sources: LaunchDarkly on Sridharan's talk, Sookocheff summary of Sridharan, TestMu AI guide, Google Testing Blog comments