Trace the OpenAI Agents SDK

Trace OpenAI Agents SDK runs, handoffs and tool calls in Python with the OpenInference instrumentor.

The OpenAI Agents SDK (openai-agents on PyPI, imported as agents) has its own tracing system. The OpenInference instrumentor converts those traces into OpenTelemetry spans, and Tracer.registerOTELInstrumentation sends them to Judgment.

pip install judgeval openai-agents openinference-instrumentation-openai-agents

from agents import Agent, Runner, function_tool
from judgeval import Tracer
from openinference.instrumentation.openai_agents import OpenAIAgentsInstrumentor

Tracer.init(project_name="lunch-order-agent")
Tracer.registerOTELInstrumentation(OpenAIAgentsInstrumentor())

@function_tool
def check_catering_capacity(restaurant: str, headcount: int) -> bool:
    """Check whether a restaurant can cater for this many people."""
    return restaurant == "Olive & Thyme"

agent = Agent(
    name="Lunch order agent",
    instructions="Help the user order lunch for their office.",
    model="gpt-5.2",
    tools=[check_catering_capacity],
)

result = Runner.run_sync(agent, "order me mediterranean food to feed 30 hungry engineers")
print(result.final_output)
Tracer.force_flush()

Each run shows up as one trace: an agent span for each agent that ran (including agents reached through a handoff), with model calls and tool calls nested under it.

Keep OpenAI’s own trace viewer

By default the instrumentor replaces the Agents SDK’s trace processors, so traces stop going to OpenAI’s dashboard. To send to both, call the instrumentor yourself with exclusive_processor=False and Judgment’s provider:

from judgeval import JudgmentTracerProvider, Tracer
from openinference.instrumentation.openai_agents import OpenAIAgentsInstrumentor

Tracer.init(project_name="lunch-order-agent")
OpenAIAgentsInstrumentor().instrument(
    tracer_provider=JudgmentTracerProvider.get_instance(),
    exclusive_processor=False,
)

If something looks wrong

If pip can’t find the package, you typed openai-agents-sdk. Install openai-agents.
If you see two spans per model call, you also wrapped an OpenAI client with wrap() or registered an OpenAI instrumentor. The Agents SDK instrumentor already records model calls.
If nothing reaches Judgment from a short script, call Tracer.force_flush() before the process exits.