The original title is: "Building AXRAY: Observability for AI Coding Agents with SigNoz and OpenTelemetry"

Short summary

The author built AXRAY, an observability tool for AI coding agents, using SigNoz and OpenTelemetry's GenAI semantic conventions. By instrumenting every agent turn with spans tagged as 'llm' or 'tool', they discovered 78% of latency came from LLM token generation, not Docker or shell commands. The article covers architecture, custom ClickHouse trace queries for per-tool latency, alert integration via SigNoz's MCP server, and deployment pitfalls.

• •AXRAY uses OpenTelemetry GenAI semantic conventions to separate LLM thinking time from tool execution time
• •78% of agent latency was traced to LLM token generation, not infrastructure—context window size was the bottleneck
• •SigNoz's ClickHouse backend enables custom SQL queries for per-tool latency; MCP server integration brings live alerts into the AXRAY UI

Generated with AI, which can make mistakes.