Available via: the Zespan tracing processor (TypeScript and Python, below) and OpenTelemetry auto-instrumentation (Python). This is not a Zespan SDK wrapper — see Send traces from any OTel-instrumented app for how the underlying endpoint works.
Zespan tracing processor
The Agents SDK lets you add a tracing processor. Zespan’s processor turns each finished span into a Zespan event, so tool calls arrive with their name, arguments, result and error — which the tool-level checks of the domain packs read.- TypeScript
- Python
The Agents SDK’s own sensitive data switch (
RunConfig.trace_include_sensitive_data in Python, traceIncludeSensitiveData in TypeScript) controls whether function spans carry input and output. When it is off, only the tool name and status are sent and the tool-level checks give “needs data”. Setting storeToolIO: false (store_tool_io=False) drops arguments and results in the same way. Set ZESPAN_OPENAI_AGENTS=off to stop the processor without a code change. Nothing in it throws into your agent.One trace per run, with agent names
- Agent names. Every span of a run carries the name of the agent it ran under, so a triage-to-billing run shows the tool calls and the final reply under
billing_agent. The hand-off span showsbilling_agentas its agent andtriage_agentas the agent it was delegated from. - The answer text. The reply is stored for both the default Responses model and the chat completions model: the model’s text, not the raw output. A turn that only calls tools, or a refusal, has no answer text, and the tools it asked for are listed on the span. Evaluators and checks read the last reply of the run as the answer.
- Times. Each span is stamped with the time it ended, so two replies in the same second are ordered correctly.
- Auto-patched model calls stay in the run. With auto-patching on, the model call a provider wrapper records joins the run’s trace under the model span, and you still get one event per model call.
Needs the new SDK releases (TypeScript and Python). With prompt storage off (
storePrompts: false, store_prompts=False) no prompt or answer text is sent.OpenTelemetry alternative
OpenAI’s Agents SDK has its own built-in tracing model (agent runs, handoffs, guardrail checks, tool calls). OpenInference’sopeninference-instrumentation-openai-agents package bridges that tracing model onto standard OpenTelemetry spans, so it flows through the same /v1/traces endpoint as everything else.
Installation
Setup
What gets captured
Agent runs, tool calls, and handoffs between agents map to Zespan’s agent/tool span kinds via the standard OpenTelemetry GenAI semantic conventions — a multi-agent handoff chain shows up as one connected trace, the same shape as Google ADK or CrewAI traces.OpenAI Agents SDK’s own model calls (to
gpt-4o etc.) are captured as part of this same instrumentation — you don’t need wrapOpenAI/patch_openai separately for agent runs built with this SDK.
