Evals are unit tests for your Agents and Teams, use them judiciously to measure and improve their performance. Agno provides 3 dimensions for evaluating Agents:

  • Accuracy: How complete/correct/accurate is the Agent’s response (LLM-as-a-judge)
  • Performance: How fast does the Agent respond and what’s the memory footprint?
  • Reliability: Does the Agent make the expected tool calls?

​
Accuracy

Accuracy evals use input/output pairs to measure your Agents and Teams performance against a gold-standard answer. Use a larger model to score the Agent’s responses (LLM-as-a-judge).

​
Example

In this example, the AccuracyEval will run the Agent with the input, then use a larger model (o4-mini) to score the Agent’s response according to the guidelines provided.

calculate_accuracy.py
from typing import Optional
from agno.agent import Agent
from agno.eval.accuracy import AccuracyEval, AccuracyResult
from agno.models.openai import OpenAIChat
from agno.tools.calculator import CalculatorTools

evaluation = AccuracyEval(
    model=OpenAIChat(id="o4-mini"),
    agent=Agent(model=OpenAIChat(id="gpt-4o"), tools=[CalculatorTools(enable_all=True)]),
    input="What is 10*5 then to the power of 2? do it step by step",
    expected_output="2500",
    additional_guidelines="Agent output should include the steps and the final answer.",
)

result: Optional[AccuracyResult] = evaluation.run(print_results=True)
assert result is not None and result.avg_score >= 8

You can also run the AccuracyEval on an existing output (without running the Agent).

accuracy_eval_with_output.py
from typing import Optional

from agno.agent import Agent
from agno.eval.accuracy import AccuracyEval, AccuracyResult
from agno.models.openai import OpenAIChat
from agno.tools.calculator import CalculatorTools

evaluation = AccuracyEval(
    model=OpenAIChat(id="o4-mini"),
    input="What is 10*5 then to the power of 2? do it step by step",
    expected_output="2500",
    num_iterations=1,
)
result_with_given_answer: Optional[AccuracyResult] = evaluation.run_with_output(
    output="2500", print_results=True
)
assert result_with_given_answer is not None and result_with_given_answer.avg_score >= 8

​
Performance

Performance evals measure the latency and memory footprint of an Agent or Team.

While latency will be dominated by the model API response time, we should still keep performance top of mind and track the Agent or Team’s performance with and without certain components. Eg: it would be good to know what’s the average latency with and without storage, memory, with a new prompt, or with a new model.

​
Example

storage_performance.py
"""Run `pip install openai agno` to install dependencies."""

from agno.agent import Agent
from agno.models.openai import OpenAIChat
from agno.eval.perf import PerfEval

def simple_response():
    agent = Agent(model=OpenAIChat(id='gpt-4o-mini'), system_message='Be concise, reply with one sentence.', add_history_to_messages=True)
    response_1 = agent.run('What is the capital of France?')
    print(response_1.content)
    response_2 = agent.run('How many people live there?')
    print(response_2.content)
    return response_2.content


simple_response_perf = PerfEval(func=simple_response, num_iterations=1, warmup_runs=0)

if __name__ == "__main__":
    simple_response_perf.run(print_results=True)

​
Reliability

What makes an Agent or Team reliable?

  • Does it make the expected tool calls?
  • Does it handle errors gracefully?
  • Does it respect the rate limits of the model API?

​
Example

The first check is to ensure the Agent makes the expected tool calls. Here’s an example:

reliability.py
from typing import Optional

from agno.agent import Agent
from agno.eval.reliability import ReliabilityEval, ReliabilityResult
from agno.tools.calculator import CalculatorTools
from agno.models.openai import OpenAIChat
from agno.run.response import RunResponse


def multiply_and_exponentiate():

    agent=Agent(
        model=OpenAIChat(id="gpt-4o-mini"),
        tools=[CalculatorTools(add=True, multiply=True, exponentiate=True)],
    )
    response: RunResponse = agent.run("What is 10*5 then to the power of 2? do it step by step")
    evaluation = ReliabilityEval(
        agent_response=response,
        expected_tool_calls=["multiply", "exponentiate"],
    )
    result: Optional[ReliabilityResult] = evaluation.run(print_results=True)
    result.assert_passed()


if __name__ == "__main__":
    multiply_and_exponentiate()

Reliability evals are currently in beta.