AI/agent pipelines quietly assume trustworthy time: stage ordering, rate limits, cache freshness, and consent/policy expiry all depend on it. When clocks drift or wander, symptoms show up elsewhere—telemetry “holes then bursts,” duplicate side effects, stale responses, or precision/recall drops in abuse detection—so teams chase model or data issues long before suspecting time. This talk presents a practical, vendor-neutral blueprint for timing integrity in production systems. We introduce a TrustedTime client contract that returns (now, confidence, source) so application code can make tiered decisions (provisional vs. final) based on time confidence. We show how to build monotonic, drift-tolerant features—epoch-based counters and windows keyed to a monotonic clock—to prevent skew-induced bucket hopping and sequence errors. We operationalize Clock-Health SLOs (offset, wander, asymmetry; plus a “confidence duty cycle”) as CI/CD and runtime gates that automatically degrade risky actions when time quality drops. An anonymized incident—lagging clock ⇒ telemetry holes ⇒ bot false positives—motivates the design. Results are illustrated with a reproducible skew-injection harness and synthetic data to demonstrate stability and sub-millisecond overhead targets. Attendees leave with SDK patterns, gating policies, and a detection/remediation playbook they can adopt immediately.