Production agentic-AI in 2026 needs four observability layers — infrastructure, LLM-call, trace, and output — and most enterprise deployments instrument only the cheaper subset (Layers 1 and 2 plus partial Layer 3); the failure modes Layers 3 and 4 catch (multi-step reasoning failure and output-distribution drift) are the ones EU AI Act Article 9 and Article 17 evidence obligations from 2 Aug 2026 onward will require coverage of, and the four layers compose directly into the four AM-110 SLA metrics (action-bounded availability, MTTD-for-Agents, output-distribution drift, per-class action error budget).

60-day cadence covers the OpenTelemetry GenAI semantic-conventions stable-promotion follow-on, the platform-consolidation arc across Datadog/New Relic AI-monitoring extensions and the Layer-4 specialists (Galileo, Arize Phoenix, Evidently), and the EU AI Act Article 9/Article 17 enforcement window opening 2 Aug 2026. Three review checks: a single-vendor four-layer coverage release reaching specialist parity at Layer 4; a standards-body publication defining agent observability primitives at the population/drift layer (CNCF OpenTelemetry GenAI working group is the candidate); a regulator enforcement action where inadequate Layer 3 or Layer 4 instrumentation was the in-scope finding.