	In Browser
	StumbleUpon
	del.icio.us
	Google
	Google Buzz
	reddit
	LinkedIn

	Facebook
	Twitter
	Linkedin
	E-Mail

Generative AI > Google Gemini API > ADK Agent Observability with OpenTelemetry

ADK Agent Observability with OpenTelemetry

Author: Venkata Sudhakar

When ADK agents call tools, invoke sub-agents, and interact with external APIs, failures are hard to debug without distributed tracing. OpenTelemetry provides vendor-neutral instrumentation that creates traces spanning the full agent lifecycle. ShopMax India uses this to diagnose latency spikes and tool failures in their multi-agent order processing pipeline.

You instrument your code by creating a tracer, wrapping key operations in spans, and exporting traces to Cloud Trace. Each agent turn, tool call, and external API request becomes a traceable span with timing and metadata. The full trace shows exactly where time was spent and where errors occurred.

The below example shows how ShopMax India adds OpenTelemetry tracing to their ADK agent tools and runners.

from opentelemetry import trace
from opentelemetry.sdk.trace import TracerProvider
from opentelemetry.sdk.trace.export import BatchSpanProcessor
from opentelemetry.exporter.cloud_trace import CloudTraceSpanExporter

# Set up Cloud Trace exporter
provider = TracerProvider()
provider.add_span_processor(
    BatchSpanProcessor(CloudTraceSpanExporter(project_id="shopmax-india"))
)
trace.set_tracer_provider(provider)
tracer = trace.get_tracer("shopmax.agents")

from google.adk.tools import FunctionTool

# Instrument a tool with tracing
def get_product_price(product_id: str) -> dict:
    with tracer.start_as_current_span("tool.get_product_price") as span:
        span.set_attribute("product_id", product_id)
        span.set_attribute("service", "shopmax-pricing")
        # Simulate pricing API call
        import time
        time.sleep(0.05)  # 50ms API call
        price = 29990  # fetched from pricing service
        span.set_attribute("price", price)
        span.set_attribute("currency", "INR")
        return {"product_id": product_id, "price": price, "currency": "INR"}

pricing_tool = FunctionTool(func=get_product_price)

Running the agent with traced tools and logging spans,

from google.adk.agents import Agent
from google.adk.runners import Runner
from google.adk.sessions import InMemorySessionService
from google.adk.types import Content, Part

agent = Agent(
    name="pricing_query_agent",
    model="gemini-2.0-flash",
    description="ShopMax India pricing query agent",
    tools=[pricing_tool]
)

session_svc = InMemorySessionService()
runner = Runner(agent=agent, session_service=session_svc, app_name="shopmax")

# Wrap the entire agent turn in a root span
with tracer.start_as_current_span("agent.turn") as root_span:
    root_span.set_attribute("user_id", "cust_5512")
    root_span.set_attribute("agent", "pricing_query_agent")

session = session_svc.create_session(app_name="shopmax", user_id="cust_5512")
    result = runner.run(
        session_id=session.id, user_id="cust_5512",
        new_message=Content(parts=[Part(text="What is the price of product SONY-WH1000XM5?")])
    )
    for event in result:
        if event.is_final_response():
            root_span.set_attribute("response_length", len(event.content.parts[0].text))
            print(event.content.parts[0].text)

print("Traces exported to Cloud Trace")

It gives the following output,

The price of Sony WH-1000XM5 is Rs 29,990.
Traces exported to Cloud Trace

# In Cloud Trace console:
Trace: agent.turn (total: 1.24s)
  |-- gemini.generate_content (0.89s)
  |-- tool.get_product_price (0.05s)
      product_id: SONY-WH1000XM5
      price: 29990

For multi-agent workflows at ShopMax India, propagate the trace context between agents using W3C TraceContext headers. This creates a single trace spanning the orchestrator agent, all sub-agents, and every tool call - making it easy to identify which agent in a 5-step pipeline caused a 3-second delay.

Send your comments, suggestions or queries regarding this site to [email protected].