In this notebook, we will learn how to monitor the internal steps (traces) of our AI agent and evaluate its performance using open-source observability tools.
The ability to observe and evaluate an agent’s behavior is essential for:
Before running this notebook, please be sure you have:
🔲 📚 Studied Introduction to Agents
🔲 📚 Studied The smolagents framework
We will need a few libraries that allow us to run, monitor, and evaluate our agents:
%pip install 'smolagents[telemetry]'
%pip install opentelemetry-sdk opentelemetry-exporter-otlp openinference-instrumentation-smolagents
%pip install langfuse datasets 'smolagents[gradio]'In this notebook, we will use Langfuse as our observability tool, but you can use any other OpenTelemetry-compatible service. The code below shows how to set environment variables for Langfuse (or any OTel endpoint) and how to instrument your smolagent.
Note: If you are using LlamaIndex or LangGraph, you can find documentation on instrumenting them here and here.
First, let’s configure the right environment variable for setting up the connection to the Langfuse OpenTelemetry endpoint.
import os
import base64
# Get your own keys from https://cloud.langfuse.com
LANGFUSE_PUBLIC_KEY = "pk-lf-..."
LANGFUSE_SECRET_KEY = "sk-lf-..."
os.environ["LANGFUSE_PUBLIC_KEY"] = LANGFUSE_PUBLIC_KEY
os.environ["LANGFUSE_SECRET_KEY"] = LANGFUSE_SECRET_KEY
os.environ["LANGFUSE_HOST"] = "https://cloud.langfuse.com" # 🇪🇺 EU region example
# os.environ["LANGFUSE_HOST"] = "https://us.cloud.langfuse.com" # 🇺🇸 US region example
LANGFUSE_AUTH = base64.b64encode(
f"{LANGFUSE_PUBLIC_KEY}:{LANGFUSE_SECRET_KEY}".encode()
).decode()
os.environ["OTEL_EXPORTER_OTLP_ENDPOINT"] = os.environ.get("LANGFUSE_HOST") + "/api/public/otel"
os.environ["OTEL_EXPORTER_OTLP_HEADERS"] = f"Authorization=Basic {LANGFUSE_AUTH}"We also need to configure our Hugging Face token for inference calls.
# Set your Hugging Face and other tokens/secrets as environment variable
os.environ["HF_TOKEN"] = "hf_..." Next, we can set up the a tracer-provider for our configured OpenTelemetry.
from opentelemetry.sdk.trace import TracerProvider
from openinference.instrumentation.smolagents import SmolagentsInstrumentor
from opentelemetry.exporter.otlp.proto.http.trace_exporter import OTLPSpanExporter
from opentelemetry.sdk.trace.export import SimpleSpanProcessor
# Create a TracerProvider for OpenTelemetry
trace_provider = TracerProvider()
# Add a SimpleSpanProcessor with the OTLPSpanExporter to send traces
trace_provider.add_span_processor(SimpleSpanProcessor(OTLPSpanExporter()))
# Set the global default tracer provider
from opentelemetry import trace
trace.set_tracer_provider(trace_provider)
tracer = trace.get_tracer(__name__)
# Instrument smolagents with the configured provider
SmolagentsInstrumentor().instrument(tracer_provider=trace_provider)Here is a simple CodeAgent from smolagents that calculates 1+1. We run it to confirm that the instrumentation is working correctly. If everything is set up correctly, you will see logs/spans in your observability dashboard.
from smolagents import HfApiModel, CodeAgent
# Create a simple agent to test instrumentation
agent = CodeAgent(
tools=[],
model=HfApiModel()
)
agent.run("1+1=")Check your Langfuse Traces Dashboard (or your chosen observability tool) to confirm that the spans and logs have been recorded.
Example screenshot from Langfuse:

Now that you have confirmed your instrumentation works, let’s try a more complex query so we can see how advanced metrics (token usage, latency, costs, etc.) are tracked.
from smolagents import (CodeAgent, DuckDuckGoSearchTool, HfApiModel)
search_tool = DuckDuckGoSearchTool()
agent = CodeAgent(tools=[search_tool], model=HfApiModel())
agent.run("How many Rubik's Cubes could you fit inside the Notre Dame Cathedral?")Most observability tools record a trace that contains spans, which represent each step of your agent’s logic. Here, the trace contains the overall agent run and sub-spans for:
You can inspect these to see precisely where time is spent, how many tokens are used, and so on:

In the previous section, we learned about the difference between online and offline evaluation. Now, we will see how to monitor your agent in production and evaluate it live.
Below, we show examples of these metrics.
Below is a screenshot showing usage for Qwen2.5-Coder-32B-Instruct calls. This is useful to see costly steps and optimize your agent.

We can also see how long it took to complete each step. In the example below, the entire conversation took 32 seconds, which you can break down by step. This helps you identify bottlenecks and optimize your agent.

You may also pass additional attributes—such as user IDs, session IDs, or tags—by setting them on the spans. For example, smolagents instrumentation uses OpenTelemetry to attach attributes like langfuse.user.id or custom tags.
from smolagents import (CodeAgent, DuckDuckGoSearchTool, HfApiModel)
from opentelemetry import trace
search_tool = DuckDuckGoSearchTool()
agent = CodeAgent(
tools=[search_tool],
model=HfApiModel()
)
with tracer.start_as_current_span("Smolagent-Trace") as span:
span.set_attribute("langfuse.user.id", "smolagent-user-123")
span.set_attribute("langfuse.session.id", "smolagent-session-123456789")
span.set_attribute("langfuse.tags", ["city-question", "testing-agents"])
agent.run("What is the capital of Germany?")
If your agent is embedded into a user interface, you can record direct user feedback (like a thumbs-up/down in a chat UI). Below is an example using Gradio to embed a chat with a simple feedback mechanism.
In the code snippet below, when a user sends a chat message, we capture the OpenTelemetry trace ID. If the user likes/dislikes the last answer, we attach a score to the trace.
import gradio as gr
from opentelemetry.trace import format_trace_id
from smolagents import (CodeAgent, HfApiModel)
from langfuse import Langfuse
langfuse = Langfuse()
model = HfApiModel()
agent = CodeAgent(tools=[], model=model, add_base_tools=True)
formatted_trace_id = None # We'll store the current trace_id globally for demonstration
def respond(prompt, history):
with trace.get_tracer(__name__).start_as_current_span("Smolagent-Trace") as span:
output = agent.run(prompt)
current_span = trace.get_current_span()
span_context = current_span.get_span_context()
trace_id = span_context.trace_id
global formatted_trace_id
formatted_trace_id = str(format_trace_id(trace_id))
langfuse.trace(id=formatted_trace_id, input=prompt, output=output)
history.append({"role": "assistant", "content": str(output)})
return history
def handle_like(data: gr.LikeData):
# For demonstration, we map user feedback to a 1 (like) or 0 (dislike)
if data.liked:
langfuse.score(
value=1,
name="user-feedback",
trace_id=formatted_trace_id
)
else:
langfuse.score(
value=0,
name="user-feedback",
trace_id=formatted_trace_id
)
with gr.Blocks() as demo:
chatbot = gr.Chatbot(label="Chat", type="messages")
prompt_box = gr.Textbox(placeholder="Type your message...", label="Your message")
# When the user presses 'Enter' on the prompt, we run 'respond'
prompt_box.submit(
fn=respond,
inputs=[prompt_box, chatbot],
outputs=chatbot
)
# When the user clicks a 'like' button on a message, we run 'handle_like'
chatbot.like(handle_like, None, None)
demo.launch()
User feedback is then captured in your observability tool:

LLM-as-a-Judge is another way to automatically evaluate your agent’s output. You can set up a separate LLM call to gauge the output’s correctness, toxicity, style, or any other criteria you care about.
Workflow:
Example from Langfuse:

# Example: Checking if the agent’s output is toxic or not.
from smolagents import (CodeAgent, DuckDuckGoSearchTool, HfApiModel)
search_tool = DuckDuckGoSearchTool()
agent = CodeAgent(tools=[search_tool], model=HfApiModel())
agent.run("Can eating carrots improve your vision?")You can see that the answer of this example is judged as “not toxic”.

All of these metrics can be visualized together in dashboards. This enables you to quickly see how your agent performs across many sessions and helps you to track quality metrics over time.

Online evaluation is essential for live feedback, but you also need offline evaluation—systematic checks before or during development. This helps maintain quality and reliability before rolling changes into production.
In offline evaluation, you typically:
Below, we demonstrate this approach with the GSM8K dataset, which contains math questions and solutions.
import pandas as pd
from datasets import load_dataset
# Fetch GSM8K from Hugging Face
dataset = load_dataset("openai/gsm8k", 'main', split='train')
df = pd.DataFrame(dataset)
print("First few rows of GSM8K dataset:")
print(df.head())Next, we create a dataset entity in Langfuse to track the runs. Then, we add each item from the dataset to the system. (If you’re not using Langfuse, you might simply store these in your own database or local file for analysis.)
from langfuse import Langfuse
langfuse = Langfuse()
langfuse_dataset_name = "gsm8k_dataset_huggingface"
# Create a dataset in Langfuse
langfuse.create_dataset(
name=langfuse_dataset_name,
description="GSM8K benchmark dataset uploaded from Huggingface",
metadata={
"date": "2025-03-10",
"type": "benchmark"
}
)for idx, row in df.iterrows():
langfuse.create_dataset_item(
dataset_name=langfuse_dataset_name,
input={"text": row["question"]},
expected_output={"text": row["answer"]},
metadata={"source_index": idx}
)
if idx >= 9: # Upload only the first 10 items for demonstration
break
We define a helper function run_smolagent() that:
Then, we loop over each dataset item, run the agent, and link the trace to the dataset item. We can also attach a quick evaluation score if desired.
from opentelemetry.trace import format_trace_id
from smolagents import (CodeAgent, HfApiModel, LiteLLMModel)
# Example: using HfApiModel or LiteLLMModel to access openai, anthropic, gemini, etc. models:
model = HfApiModel()
agent = CodeAgent(
tools=[],
model=model,
add_base_tools=True
)
def run_smolagent(question):
with tracer.start_as_current_span("Smolagent-Trace") as span:
span.set_attribute("langfuse.tag", "dataset-run")
output = agent.run(question)
current_span = trace.get_current_span()
span_context = current_span.get_span_context()
trace_id = span_context.trace_id
formatted_trace_id = format_trace_id(trace_id)
langfuse_trace = langfuse.trace(
id=formatted_trace_id,
input=question,
output=output
)
return langfuse_trace, outputdataset = langfuse.get_dataset(langfuse_dataset_name)
# Run our agent against each dataset item (limited to first 10 above)
for item in dataset.items:
langfuse_trace, output = run_smolagent(item.input["text"])
# Link the trace to the dataset item for analysis
item.link(
langfuse_trace,
run_name="smolagent-notebook-run-01",
run_metadata={ "model": model.model_id }
)
# Optionally, store a quick evaluation score for demonstration
langfuse_trace.score(
name="<example_eval>",
value=1,
comment="This is a comment"
)
# Flush data to ensure all telemetry is sent
langfuse.flush()You can repeat this process with different:
Then compare them side-by-side in your observability tool:

In this notebook, we covered how to:
🤗 Happy coding!
< > Update on GitHub