pytracer-autogen

pyTracer adapter for AutoGen (0.4+). AutoGen instruments agents, tool calls, and model calls with OpenTelemetry using the GenAI semantic conventions. This adapter plugs a pyTracer exporter into an OpenTelemetry tracer provider, so those spans are captured as pyTracer spans with no @op in the agent code.

AutoGen emits its GenAI model and tool spans through the global OpenTelemetry provider, so register pyTracer globally with install() before building the runtime. Configure pyTracer first, so the exporter records into your transport:

import pytracer_sdk as pytracer
from pytracer_autogen import install
from autogen_core import SingleThreadedAgentRuntime

pytracer.configure(
    pytracer.HTTPTransport("https://gateway.internal", api_key="YOUR_KEY"),
    service_name="my-agent", branch="main", commit="abc123",
)

install()  # register pyTracer as the global OpenTelemetry provider
runtime = SingleThreadedAgentRuntime()

Passing a provider only to the runtime (SingleThreadedAgentRuntime(tracer_provider=pytracer_tracer_provider())) sees just the runtime's internal messaging spans, which the exporter drops as noise, so install() is the setup that captures the actual model and tool calls.

Because it keys off standard GenAI attributes (gen_ai.operation.name, gen_ai.tool.name, gen_ai.agent.name, gen_ai.usage.*), the same exporter captures spans from any GenAI-instrumented library, not only AutoGen.

Versions: what is captured depends on the AutoGen version emitting GenAI spans. Some versions emit no model-call span or tool result; pyTracer records what the framework reports, so upgrade AutoGen (or add the relevant GenAI instrumentation) if model calls or tool outputs are missing.

Replay

This adapter captures runs. Replaying a captured run currently requires @op instrumentation, because the replay engine intercepts @op calls; replay through the adapter itself is on the roadmap. Attempting to replay an adapter-only trace is flagged (the run has recordings but hits no @op seam) rather than silently running live. See the main README for the replay workflow.