# Agent Observability and Tracing
Status: public
Confidence: medium (0.865) (verified)
Last verified: 2026-06-02
Generation: ai_structured


## TL;DR

Agent observability is the practice of recording what an agent did, which model calls it made, which tools it invoked, which guardrails or handoffs ran, and how each step contributed to the final answer.

## Core Explanation

Tracing turns an agent run into a structured execution record. For agent engineering, this is useful for debugging tool failures, measuring latency and cost, reviewing safety decisions, reproducing bad answers, and comparing evaluation runs. OpenTelemetry provides a general trace model, while agent and GenAI conventions add AI-specific span metadata.

## Source-Mapped Facts

- OpenAI Agents SDK tracing documentation says tracing collects a record of events during an agent run, including LLM generations, tool calls, handoffs, guardrails, and custom events. ([source](https://openai.github.io/openai-agents-python/tracing/))
- OpenTelemetry describes traces as records of the paths taken by requests as they propagate through multi-service architectures. ([source](https://opentelemetry.io/docs/concepts/signals/traces/))
- OpenTelemetry generative AI semantic conventions define span conventions for generative AI client operations. ([source](https://opentelemetry.io/docs/specs/semconv/gen-ai/gen-ai-spans/))

## Further Reading

- [OpenAI Agents SDK tracing](https://openai.github.io/openai-agents-python/tracing/)
- [OpenTelemetry traces](https://opentelemetry.io/docs/concepts/signals/traces/)
- [OpenTelemetry GenAI spans](https://opentelemetry.io/docs/specs/semconv/gen-ai/gen-ai-spans/)