Enterprise multi-agent AI systems produce thousands of inter-agent interactions per hour, yet existing observability tools capture these dependencies without enforcing anything. OpenTelemetry and Langfuse collect telemetry but treat governance as a downstream analytics concern, not a real-time enforcement target. The result is an “observe-but-do-not-act” gap where policy violations are detected only after damage is done. We present Governance-Aware Agent Telemetry (GAAT), a reference architecture that closes the loop between telemetry collection and automated policy enforcement for multi-agent systems. GAAT introduces (1) a Governance Telemetry Schema (GTS) extending OpenTelemetry with governance attributes; (2) a real-time policy violation detection engine using OPA-compatible declarative rules under sub-200 ms latency; (3) a Governance Enforcement Bus (GEB) with graduated interventions; and (4) a Trusted Telemetry Plane with cryptographic provenance. We evaluated GAAT against four baseline systems across data residency, bias detection, authorization compliance, and adversarial telemetry scenarios. On a live five-agent e-commerce system, GAAT achieved 98.3% Violation Prevention Rate (VPR, ±0.7%) on 5,000 synthetic injection flows across 10 independent runs, with 8.4 ms median detection latency and 127 ms median end-to-end enforcement latency. On 12,000 empirical production-realistic traces, GAAT achieved 99.7% VPR; residual failures (∼40% timing edge cases, ∼35% ambiguous PII classification, ∼25% incomplete lineage chains). Statistical validation confirmed significance with 95% bootstrap confidence intervals [97.1%, 99.2%] (p < 0.001 vs all baselines). GAAT outperformed NeMo Guardrails-style agent-boundary enforcement by 19.5 percentage points (78.8% VPR vs 98.3%). We also provide formal property specifications for escalation termination, conflict resolution determinism, and bounded false quarantine—each with explicit assumptions—validated through 10,000 Monte Carlo simulations.