openai-agents-opentelemetry

OpenTelemetry tracing processor for the OpenAI Agents SDK.

Bridges agent traces to any OTLP-compatible backend (Jaeger, Datadog, Honeycomb, Grafana Tempo, Langfuse, etc.).

Installation

pip install openai-agents-opentelemetry

Quick Start

from agents import Agent, Runner, add_trace_processor
from openai_agents_opentelemetry import OpenTelemetryTracingProcessor

# Create and register the OpenTelemetry processor
otel_processor = OpenTelemetryTracingProcessor()
add_trace_processor(otel_processor)

# Now all agent traces will be exported to your configured OTel backend
agent = Agent(name="Assistant", instructions="You are helpful.")
result = await Runner.run(agent, "Hello!")

Using OpenTelemetry Only (No OpenAI Backend)

If you want traces to go only to your OpenTelemetry backend:

from agents import set_trace_processors
from openai_agents_opentelemetry import OpenTelemetryTracingProcessor

# Replace the default processor entirely
otel_processor = OpenTelemetryTracingProcessor()
set_trace_processors([otel_processor])

Span Mapping

The processor maps SDK spans to OpenTelemetry spans following OpenTelemetry Semantic Conventions for GenAI:

Configuration

Content Capture Configuration

Control what content is captured for privacy and compliance using ProcessorConfig:

from openai_agents_opentelemetry import OpenTelemetryTracingProcessor, ProcessorConfig

config = ProcessorConfig(
    capture_prompts=True,        # Capture prompt content as span events
    capture_completions=True,    # Capture completion content as span events
    capture_tool_inputs=True,    # Capture tool input arguments
    capture_tool_outputs=True,   # Capture tool output results
    max_attribute_length=4096,   # Max length for span attributes
    max_event_length=8192,       # Max length for span event attributes
    baggage_keys=["user.id", "session.id"],  # Propagate context to spans
)

processor = OpenTelemetryTracingProcessor(config=config)

Content Filtering / PII Redaction

Apply custom filtering to redact sensitive data before capture:

import re
from openai_agents_opentelemetry import OpenTelemetryTracingProcessor, ProcessorConfig

def redact_pii(content: str, context: str) -> str:
    """Custom PII redaction callback."""
    # Redact SSNs
    content = re.sub(r"\b\d{3}-\d{2}-\d{4}\b", "[SSN REDACTED]", content)
    # Redact email addresses
    content = re.sub(r"\b[\w.-]+@[\w.-]+\.\w+\b", "[EMAIL REDACTED]", content)
    return content

config = ProcessorConfig(
    capture_prompts=True,
    capture_completions=True,
    content_filter=redact_pii,
)

processor = OpenTelemetryTracingProcessor(config=config)

OpenTelemetry SDK Configuration

The processor uses the globally configured OpenTelemetry TracerProvider. Configure it as you normally would:

from opentelemetry import trace
from opentelemetry.sdk.trace import TracerProvider
from opentelemetry.sdk.trace.export import BatchSpanProcessor
from opentelemetry.exporter.otlp.proto.grpc.trace_exporter import OTLPSpanExporter

# Configure OpenTelemetry
provider = TracerProvider()
processor = BatchSpanProcessor(OTLPSpanExporter(endpoint="http://localhost:4317"))
provider.add_span_processor(processor)
trace.set_tracer_provider(provider)

# Then add the Agents SDK processor
from agents import add_trace_processor
from openai_agents_opentelemetry import OpenTelemetryTracingProcessor

add_trace_processor(OpenTelemetryTracingProcessor())

Resource Helper

Use the create_resource helper for standard resource attributes:

from opentelemetry import trace
from opentelemetry.sdk.trace import TracerProvider
from openai_agents_opentelemetry import create_resource

resource = create_resource(
    service_name="my-agent-service",
    service_version="1.0.0",
    additional_attributes={"deployment.environment": "production"},
)

provider = TracerProvider(resource=resource)
trace.set_tracer_provider(provider)

Span Events

The processor automatically adds span events for content capture (controlled by ProcessorConfig):

Metrics

Enable metrics collection for token usage, operation duration, and agent-specific counters:

from openai_agents_opentelemetry import OpenTelemetryTracingProcessor

# Enable metrics collection
processor = OpenTelemetryTracingProcessor(enable_metrics=True)

Standard OTel GenAI Metrics

Agent-Specific Metrics

Configuring Metrics with OpenTelemetry SDK

Use create_metrics_views() to configure histogram bucket boundaries according to the OpenTelemetry Semantic Conventions for GenAI. Without these views, the SDK uses default buckets that are not suitable for GenAI workloads:

• Token counts can range from 1 to millions (large context windows like GPT-4 128k or Claude 200k)
• Operation durations follow different patterns than typical HTTP requests

from opentelemetry import metrics
from opentelemetry.sdk.metrics import MeterProvider
from opentelemetry.sdk.metrics.export import PeriodicExportingMetricReader
from opentelemetry.exporter.otlp.proto.grpc.metric_exporter import OTLPMetricExporter
from openai_agents_opentelemetry import (
    OpenTelemetryTracingProcessor,
    create_metrics_views,
)

# Configure metrics with recommended GenAI histogram buckets
exporter = OTLPMetricExporter(endpoint="http://localhost:4317")
reader = PeriodicExportingMetricReader(exporter)
views = create_metrics_views()  # Returns views with OTel GenAI recommended bucket boundaries
provider = MeterProvider(metric_readers=[reader], views=views)
metrics.set_meter_provider(provider)

# Then enable metrics in the processor
from agents import add_trace_processor

add_trace_processor(OpenTelemetryTracingProcessor(enable_metrics=True))

The create_metrics_views() function configures:

• gen_ai.client.token.usage: Buckets [1, 4, 16, 64, 256, 1024, 4096, 16384, 65536, 262144, 1048576, 4194304, 16777216, 67108864]
• gen_ai.client.operation.duration: Buckets [0.01, 0.02, 0.04, 0.08, 0.16, 0.32, 0.64, 1.28, 2.56, 5.12, 10.24, 20.48, 40.96, 81.92]

You can also access these bucket constants directly if needed:

from openai_agents_opentelemetry import TOKEN_BUCKETS, DURATION_BUCKETS

Context Propagation (Baggage)

Propagate context like user IDs or session IDs across all agent spans using OpenTelemetry baggage:

from opentelemetry import baggage, context
from openai_agents_opentelemetry import OpenTelemetryTracingProcessor, ProcessorConfig

# Configure which baggage keys to read and add as span attributes
config = ProcessorConfig(
    baggage_keys=["user.id", "session.id", "tenant.id"]
)
processor = OpenTelemetryTracingProcessor(config=config)

# Set baggage before running agents
ctx = baggage.set_baggage("user.id", "user-123")
ctx = baggage.set_baggage("session.id", "session-456", context=ctx)

with context.attach(ctx):
    # All spans created during this agent run will have user.id and session.id attributes
    result = await Runner.run(agent, "Hello!")

This enables filtering traces by user, session, or tenant in your observability backend.

Compatibility

This package is tested weekly against the latest OpenAI Agents SDK to ensure compatibility.

License

MIT