Traditional MLOps monitoring assumes stateless inference with comparable outputs across runs, single decision boundaries, timely ground truth, and human oversight. These assumptions break when agents loop and call tools. Most teams migrate by addition—layering new agent spans onto existing stacks—leaving drift monitors and retraining triggers unchanged.

The result: green traces on failed runs. OpenTelemetry's Gen AI semantic conventions define create_agent, invoke_agent, execute_tool, and plan spans, with Langfuse, LangSmith, Arize Phoenix, W&B Weave, and Agent Ops emitting versions. However, inherited signals still fire falsely healthy.

Comparable outputs fail when identical inputs yield different results within days. Stateless inference breaks when early missteps cascade through loops. Single decision boundaries disappear in agentic systems.

Ground truth delays and human-in-the-loop assumptions collapse. Tau-bench reveals pass^k reliability: gpt-4o clears 61% of retail tasks once, but 8 attempts yield under 25% success. Monitoring stacks fail by staying green while masking systemic issues.