What an AI Agent Observability Standard Actually Solves
Traditional application observability tools trace requests, services, and database calls using well-established span models. AI agents introduce a different shape of problem: a single user request can trigger multiple model invocations, tool calls, retrieval steps, and sub-agent delegations, each with its own inputs, outputs, and decision context. Standard APM tooling was not designed to represent that structure, which is why a layer of AI-specific semantic conventions has emerged on top of OpenTelemetry.
An AI agent observability standard defines how to represent these events as spans, what attributes those spans should carry, and how parent-child relationships should be structured so that a full agent execution chain can be reconstructed after the fact. OpenInference and OpenLLMetry are the two most referenced open-source efforts in this space. Both are built on OpenTelemetry rather than replacing it, which means teams with existing OTel pipelines can generally adopt either without re-architecting their telemetry infrastructure.
Shared Foundation: Why Both Standards Build on OpenTelemetry
Neither OpenInference nor OpenLLMetry is a formally ratified standard. Both are community and vendor-led open-source specifications layered on OpenTelemetry, the CNCF project that provides vendor-neutral APIs, SDKs, and the OTLP wire protocol for traces, metrics, and logs. This matters for enterprise adoption because it means traces produced by either framework can be routed to existing OpenTelemetry-compatible backends such as Jaeger, Grafana Tempo, or Honeycomb without introducing a new, incompatible telemetry pipeline.
OpenTelemetry is also separately developing its own official Semantic Conventions for Generative AI systems, defining standardized attributes for LLM request and response spans. OpenInference, OpenLLMetry, and this in-progress OTel GenAI effort do not yet fully converge, which creates a practical risk: attribute naming and span-kind definitions can diverge across the three, requiring normalization at the backend if an organization mixes instrumentation sources.