Standard
Draft OpenTelemetry conventions defining span types for model inference, tool execution and retrieval, with attributes like model IDs and token usage.
MCP's 2026-07-28 spec deprecated protocol-level Logging and left it at least twelve months of support. A dev.to guide says OpenTelemetry's MCP conventions replace it, and since they are still unstable, teams should pin instrumentation versions.
Reality
LangChain's survey of 1,340 practitioners found 89% had observability on their agents and 37.3% ran online evaluations. The OpenTelemetry GenAI attribute registry explains why only the second number measures quality.
Reality
Microsoft Foundry writes agent runs as OpenTelemetry spans with gen_ai attributes and exports them to Azure Monitor, where evaluators score sampled production traffic off the same records. Those attributes include tool arguments.
Reality
Kept's tracing layer went in before the agent loop, which produced one design decision worth copying and a promised accounting of the drawbacks that the available text never reaches.
Reality
Bedrock AgentCore Evaluations promises to score agents built on LangGraph, LlamaIndex, OpenAI's SDK, Google ADK or the Claude Agent SDK. The catch sits in the instrumentation.
Reality
The New Stack argues an agent run record is application data with a telemetry-shaped workload. Flatten one reported event rate and the daily average hides a burst about 19 times larger.
Reality