Skip to main content

Compatibility Matrix

Last updated on

Trace attribution needs OpenTelemetry traces with GenAI semantic conventions. The tables below show which model SDKs, gateways, and frameworks emit those attributes natively, and which are supported by the Harness instrumentation SDK.

Use this matrix to decide your setup path: if your stack already emits GenAI traces, route them to Harness; if not, add instrumentation. Go to AI Cost Management Quickstart for the steps.

Model Client SDKs

Used for LLM inference.

SDKGenAI semconv export?Support in Harness instrumentation SDK
OpenAI Python SDK✅ Yes (most mature)✅ Yes
Anthropic SDK✅ Yes✅ Yes
Google GenAI / Vertex SDK✅ Yes✅ Yes
AWS Bedrock✅ YesComing soon

Gateways

GatewayGenAI semconv export?Support in Harness instrumentation SDK
LiteLLM✅ Yes✅ Yes
Envoy❌ NoN/A

Agent Orchestration Frameworks

FrameworkGenAI semconv export?
Google Agent Development Kit✅ Native
OpenAI Agents SDK ("OpenAI ADK")✅ Yes
Anthropic Agents SDK✅ Yes
Microsoft Agent Framework ("Microsoft ADK")✅ Native (partial)
Microsoft Semantic Kernel✅ Native (opt-in)
AWS Strands Agents SDK✅ Native
LangGraph / LangChain⚠️ Not native
CrewAI⚠️ Mixed
LlamaIndex⚠️ Not native
Pydantic AI✅ Native
AutoGen / AG2⚠️ Partial
Haystack⚠️ Not native
Smolagents⚠️ Not native
Frameworks marked "Not native" still work

A framework that does not emit GenAI semantic conventions natively can still be instrumented. Use the OpenInference instrumentation or the LangSmith OTel export for LangChain and LangGraph, or the Harness SDK when the framework routes through LiteLLM. Go to Set Up AI Cost Traces for per-framework steps.


Next Steps