What is Agent Observability?

Short summary

Agent observability captures an AI agent's complete decision-making journey—every LLM call, tool invocation, token cost, and reasoning step—at production-grade detail to debug and audit behavior. Unlike service observability, agent systems require tracing nondeterminism, deeply nested tool calls, and prompts as data. Traces, metrics, and logs form the three pillars, with OpenTelemetry GenAI semantic conventions emerging as the platform standard.

• •Agent observability records the full execution path: LLM calls, tool invocations, token usage, latency, and reasoning chains
• •Agents require different observability than services due to nondeterminism, nested calls, and prompts as data artifacts
• •OpenTelemetry GenAI semantic conventions are becoming the standard for capturing agent telemetry across tools and platforms

Generated with AI, which can make mistakes.