OpenTelemetry makes GenAI observability a shared contract
OpenTelemetry's GenAI semantic conventions give agent builders a common language for instrumenting model calls, token usage, operation names, request parameters, responses, and error paths.
Why this signal matters
Agent products cannot be operated safely if model calls, tool hops, token consumption, and failure modes disappear into vendor-specific logs. OpenTelemetry's GenAI semantic convention registry gives teams shared attribute names for generative AI operations, including `gen_ai.operation.name`, `gen_ai.provider.name`, `gen_ai.request.model`, `gen_ai.response.model`, `gen_ai.tool.call.*`, and `gen_ai.usage.input_tokens` / `gen_ai.usage.output_tokens`. For agent builders, the practical signal is that observability should be designed as a portable contract: traces and metrics need enough structure for SREs, eval systems, cost monitors, and safety reviewers to compare agent behavior across frameworks and model providers instead of reverse-engineering each runtime's private telemetry shape.
Actionable summary
When reviewing agent observability, inspect OpenTelemetry GenAI attributes for operation names, system/model identity, request and response metadata, token accounting, errors, and trace correlation.
- Agent usefulness
- 91/100
- Confidence
- 82%
- Canonical data
- JSON + Markdown