The OpenTelemetry project published OpenTelemetry for Generative AI, introducing standardized gen_ai.* semantic conventions for model names, token usage histograms, and opt-in prompt/completion events across LLM frameworks.
Observability & Tracing
Also known as: LLM Observability · OpenTelemetry GenAI · Agent Tracing · Langfuse
Plain English
Recording every prompt, tool call, retrieval step, latency spike, and token cost inside an AI app or agent so engineers can debug failures and measure quality in production.
Technical Overview
AI observability instruments multi-step agent loops, RAG retrievals, and tool invocations as nested trace spans carrying model IDs, token counts, and latency metrics. OpenTelemetry's gen_ai.* semantic conventions provide a vendor-neutral OTLP schema adopted across frameworks and open-source tracing platforms like Langfuse and Arize Phoenix, feeding directly into online evaluations.
Timeline (2)
2024
- Ecosystem
2026
- Ecosystem
OpenTelemetry expanded its GenAI observability specifications for multi-agent workflows and MCP tool calls, unifying trace propagation across distributed agent harnesses and inference providers.
See also (bkz.) (6)
Open full graphObservability & Tracing