Why should you care?

Slow CRM answers need span breakdown — was latency embedding, Pinecone, or GPT?

See it live — copy this example

Paste into an ASP.NET Core 8+ / AgentNest project, then run with dotnet run (set your API keys in user-secrets).

// AgentNest.Telemetry/AgentTelemetry.cs
public sealed class AgentTelemetry
{
    static readonly ActivitySource Source = new("AgentNest.Agents");
    public Activity Start(string agentName) =>
        Source.StartActivity(agentName, ActivityKind.Internal);
}

// In RagPipeline
using var span = telemetry.Start("rag.ask");
span?.SetTag("tenant.id", tenantId);
span?.SetTag("rag.topk", topK);

What happened?

• ActivitySource creates spans exported to Azure Monitor Application Insights.
• Tags attach tenant and RAG parameters for trace search.

Practice next

1. Add OpenTelemetry.AspNetCore and Azure Monitor exporter.
2. Instrument AgentBase, RagPipeline, and kernel InvokeAsync.
3. Propagate trace context to Service Bus messages.
4. Add span events on each tool invocation with function name.
5. Sample traces at 10% for high-volume CRM tenants.

Remember

OpenTelemetry traces tie HTTP to agent, RAG, and tool spans. Tag spans with tenant and agent metadata. Export to Application Insights for AgentNest ops.

Slow RAG investigation

ERP copilot p95 jumps from 2s to 9s.

Outcome: Trace shows Milvus search span grew — HNSW index rebuild fixed it.