OpenTelemetry GenAI LLM Agent Tracing Sandbox PoC

Goal

Create a small local proof that an LLM-agent run can be represented with OpenTelemetry GenAI-style spans:

one parent invoke_agent span
one chat LLM client span
two execute_tool child spans
model, provider, token-count, finish-reason, and tool-call attributes

No production credentials, API keys, real user prompts, or live LLM calls were used.

Commands Run

python3 --version
docker --version
python3 -m pip index versions opentelemetry-sdk
python3 -m pip index versions opentelemetry-instrumentation-anthropic

Relevant output:

Python 3.12.8
Docker version 29.2.0, build 0b9d198
opentelemetry-sdk (1.42.1)
opentelemetry-instrumentation-anthropic (0.61.0)

Sandbox setup:

rm -rf /tmp/effloow-otel-genai-poc
mkdir -p /tmp/effloow-otel-genai-poc
python3 -m venv /tmp/effloow-otel-genai-poc/.venv
/tmp/effloow-otel-genai-poc/.venv/bin/python -m pip install --upgrade pip
/tmp/effloow-otel-genai-poc/.venv/bin/python -m pip install \
  opentelemetry-sdk==1.42.1 \
  opentelemetry-exporter-otlp==1.42.1 \
  opentelemetry-instrumentation-anthropic==0.61.0

Installed package output:

opentelemetry-api==1.42.1
opentelemetry-exporter-otlp==1.42.1
opentelemetry-instrumentation-anthropic==0.61.0
opentelemetry-instrumentation==0.63b1
opentelemetry-proto==1.42.1
opentelemetry-sdk==1.42.1
opentelemetry-semantic-conventions-ai==0.5.1
opentelemetry-semantic-conventions==0.63b1

Failure Found

Importing opentelemetry.instrumentation.anthropic failed after installing only the instrumentation package:

importlib.metadata.PackageNotFoundError: No package metadata was found for pydantic

The instrumentation package imported successfully after adding the Anthropic client and Pydantic:

/tmp/effloow-otel-genai-poc/.venv/bin/python -m pip install anthropic pydantic

Relevant installed versions:

anthropic==0.105.2
pydantic==2.13.4
opentelemetry-instrumentation-anthropic==0.61.0

Trace Export

The PoC created a custom JSON span exporter and manually emitted the agent span tree. The trace was written to:

/tmp/effloow-otel-genai-poc/trace-export.json

Summary output:

/tmp/effloow-otel-genai-poc/trace-export.json
span_count 4
chat CLIENT 211221f62b7b1a6e ['gen_ai.operation.name', 'gen_ai.provider.name', 'gen_ai.request.model', 'gen_ai.response.finish_reasons', 'gen_ai.response.model', 'gen_ai.system', 'gen_ai.usage.input_tokens', 'gen_ai.usage.output_tokens']
execute_tool INTERNAL 211221f62b7b1a6e ['effloow.output.rows', 'gen_ai.operation.name', 'gen_ai.tool.call.id', 'gen_ai.tool.name']
execute_tool INTERNAL 211221f62b7b1a6e ['effloow.output.rows', 'gen_ai.operation.name', 'gen_ai.tool.call.id', 'gen_ai.tool.name']
invoke_agent INTERNAL None ['effloow.lab_run.slug', 'gen_ai.agent.name', 'gen_ai.operation.name']

Trace JSON summary:

{
  "environment": {
    "python": "3.12.8",
    "platform": "macOS-15.6-arm64-arm-64bit",
    "otel_sdk": "1.42.1",
    "otel_exporter_otlp": "1.42.1",
    "anthropic_instrumentation": "0.61.0",
    "anthropic_client": "0.105.2",
    "api_calls_made": false
  },
  "span_count": 4,
  "span_names": ["chat", "execute_tool", "execute_tool", "invoke_agent"],
  "trace_id": "0f1035558bef566e0d26981c0031d202"
}

Jaeger Attempt

Command attempted:

docker run -d --name effloow-otel-jaeger-poc \
  -e COLLECTOR_OTLP_ENABLED=true \
  -p 16686:16686 \
  -p 4318:4318 \
  jaegertracing/all-in-one:latest

Result:

Unable to find image 'jaegertracing/all-in-one:latest' locally
docker: error getting credentials - err: signal: terminated, out: ``

Docker blocked in credential lookup while pulling the image, so the process was terminated. No Jaeger UI screenshot was captured, and no claim should be made that this run verified the Jaeger UI.

What Worked

Python OpenTelemetry SDK 1.42.1 installed cleanly in a temporary virtualenv.
Anthropic instrumentation 0.61.0 installed, but required the Anthropic client/Pydantic path to import successfully in this sandbox.
A local trace export captured a coherent agent tree with one root invoke_agent, one chat span, and two execute_tool spans.
The trace included both legacy-style gen_ai.system and newer gen_ai.provider.name provider attributes, making the example resilient to convention transition.
Prompt and tool payload content was intentionally not recorded.

What Failed or Was Not Tested

No live Claude or OpenAI request was made.
No API key was loaded or used.
Automatic Anthropic client instrumentation was imported but not exercised against a real Anthropic SDK call.
Jaeger all-in-one did not start because Docker image pull credential lookup blocked.
No OTLP backend ingest, Jaeger UI screenshot, production collector config, sampling policy, cost dashboard, or eval metric was verified.

Limitations for the Article

The article may truthfully say that Effloow Lab ran a local OpenTelemetry span-export PoC for an agent-shaped trace. It must not say that Effloow verified live Anthropic/OpenAI auto-instrumentation, Jaeger UI rendering, provider token accounting, or production observability behavior.