MCP Meets OpenTelemetry: Bridging Agent Observability and Infrastructure Monitoring

There are two worlds in production observability right now, and they do not talk to each other. The first world is infrastructure monitoring. Prometheus scrapes metrics. OpenTelemetry collectors ship traces and logs to Datadog, Grafana Tempo, Jaeger. Your SRE team has dashboards for p99 latency, error rates, throughput. This stack is mature. It works. Teams have spent years building runbooks around it. The second world is agent observability. What did the LLM actually do? Did it hallucinate? Did it drop context? How much did this execution cost? What was the eval score? These questions live in a completely separate tool -- a different dashboard, a different data model, a different team. I have been building Iris, an MCP-native agent eval and observability tool, for the past several months. The more I work with production agent deployments, the more convinced I am that these two worlds need to merge. Not eventually. Now. The Two-Dashboard Problem Here is a scenario I have seen play out