Maple agent tracing: Microsoft Agent Framework and Semantic Kernel

Goal: one user conversation = one Maple Agent Session, with a transcript, every model call, every tool call (arguments, results, failures) and tokens.

The framework emits the spans itself. You add: an OTLP/HTTP exporter, content capture, a gen_ai.conversation.id span processor (the framework never sets one for local-history sessions), and a flush.

Step 0: Detect

• Python MAF: agent-framework, agent-framework-core in pyproject.toml / requirements*.txt / uv.lock. Check the installed version (python -c "import agent_framework; print(agent_framework.__version__)"). Target ≥ 1.19.0; upgrade if older (1.13 lacks the otlp_* arguments used below).
• .NET MAF: Microsoft.Agents.AI in *.csproj. Target ≥ 1.22.0.
• Semantic Kernel: semantic-kernel (Python, target ≥ 1.44.1) or Microsoft.SemanticKernel (.NET). Use the Semantic Kernel section.
• Existing OpenTelemetry: search for TracerProvider(, set_tracer_provider, configure_azure_monitor, logfire.configure, configure_otel_providers, AddOpenTelemetry(, Sdk.CreateTracerProviderBuilder. If a provider exists, add Maple's exporter and the conversation processor to it; do not create a second provider and do not call configure_otel_providers().
• Find every place a user message is handled (HTTP route, queue consumer, CLI loop) and what identifies the conversation there (chat id, thread id, AgentSession). You need it in Step 3.

Step 1: Key and region