I Traced a Multi-Step LLM Agent With Self-Hosted SigNoz. One Feature Sold Me.

Short summary

A technical walkthrough of self-hosting SigNoz to trace a four-step LLM agent pipeline using OpenTelemetry GenAI semantic conventions. The author demonstrates how trace spans with gen_ai attributes reveal model selection, token usage, and cost per call. The standout feature turned out to be the flame graph waterfall that maps directly to the agent's mental model, making debugging expensive or slow agent runs visual and concrete.

• •Self-hosted SigNoz with OpenTelemetry traces a 4-step LLM agent pipeline effectively
• •GenAI semantic conventions on spans expose model, token counts, and cost per LLM call
• •Flame graph waterfall maps directly to agent execution flow, making debugging visual

Generated with AI, which can make mistakes.