Observability for Agents: Traces, Spans, and Failure Debugging in Agentic AI
Observability for Agents: Traces, Spans, and Failure Debugging
What to capture
- Run ID
- Prompts and versions
- Tool calls and payloads
- Retrieved context
- Errors and retries
Spans
Break runs into spans (plan, tool call, summarize, verify). This helps identify where latency and failures happen.

