How OpenTelemetry Tracing Can Debug Complex Multi-Step AI Agent Runs

Debugging multi-step AI agent failures requires more than logging the first and last request — every intermediate decision must be traceable. OpenTelemetry's GenAI semantic conventions provide standardised field names for agent spans, making traces compatible with existing and future tooling without inventing a custom schema. A recommended trace structure nests tool calls, compaction events, and model interactions as child spans under each step, enabling queries like p95 step latency by tool without additional engineering work. Key attributes to capture include tokens per step, cumulative run cost, halt reason, content-hashed message payloads stored by reference, and retry or provider metadata. Together, these attributes allow teams to identify exactly where an agent went wrong, why a tool was called repeatedly, and how costs accumulated across a full run.