How to Intermediate · 3 min read

How to evaluate AI agent performance

Quick answer

Evaluate AI agent performance by measuring key metrics such as task success rate, response relevance, latency, and user satisfaction using automated tests or human evaluation. Use prompt engineering and benchmark datasets to systematically assess agent capabilities and robustness.

PREREQUISITES

Python 3.8+
OpenAI API key (free tier works)
pip install openai>=1.0

Setup

Install the OpenAI Python SDK and set your API key as an environment variable to interact with AI agents programmatically.

bash

pip install openai

Step by step

Use the OpenAI API to send tasks to your AI agent and evaluate its responses against expected outputs. Measure metrics like success rate, response time, and qualitative scores.

python

import os
from openai import OpenAI
import time

client = OpenAI(api_key=os.environ["OPENAI_API_KEY"])

# Define test cases with input prompts and expected outputs
test_cases = [
    {"input": "Translate 'Hello' to French.", "expected": "Bonjour"},
    {"input": "What is 5 plus 7?", "expected": "12"},
    {"input": "Summarize the plot of 'The Matrix'.", "expected": None}  # subjective
]

success_count = 0
latencies = []

for case in test_cases:
    start = time.time()
    response = client.chat.completions.create(
        model="gpt-4o",
        messages=[{"role": "user", "content": case["input"]}]
    )
    latency = time.time() - start
    latencies.append(latency)
    answer = response.choices[0].message.content.strip()

    # Simple exact match for objective tasks
    if case["expected"]:
        if answer.lower() == case["expected"].lower():
            success_count += 1
    else:
        # For subjective tasks, manual or heuristic evaluation needed
        print(f"Input: {case['input']}\nAgent answer: {answer}\n")

print(f"Success rate: {success_count}/{len(test_cases)}")
print(f"Average latency: {sum(latencies)/len(latencies):.2f} seconds")

output

Input: Summarize the plot of 'The Matrix'.
Agent answer: The Matrix is a sci-fi film where a hacker discovers reality is a simulation controlled by machines.

Success rate: 2/3
Average latency: 1.23 seconds

Common variations

You can evaluate asynchronously, use streaming responses for latency measurement, or switch models like claude-3-5-sonnet-20241022 for better coding or reasoning tasks. Human evaluation can complement automated metrics for subjective tasks.

python

import asyncio
import os
from openai import OpenAI

client = OpenAI(api_key=os.environ["OPENAI_API_KEY"])

async def evaluate_async(prompt):
    response = await client.chat.completions.acreate(
        model="gpt-4o",
        messages=[{"role": "user", "content": prompt}]
    )
    return response.choices[0].message.content

async def main():
    prompts = ["Explain quantum computing.", "Write a Python function to reverse a string."]
    results = await asyncio.gather(*(evaluate_async(p) for p in prompts))
    for prompt, answer in zip(prompts, results):
        print(f"Prompt: {prompt}\nAnswer: {answer}\n")

asyncio.run(main())

output

Prompt: Explain quantum computing.
Answer: Quantum computing uses quantum bits that can be in multiple states simultaneously, enabling complex computations.

Prompt: Write a Python function to reverse a string.
Answer: def reverse_string(s):\n    return s[::-1]

Troubleshooting

If you see inconsistent results, ensure your prompts are clear and unambiguous. For latency spikes, check your network and API usage limits. Use logging to track agent responses and errors for debugging.

✅

Key Takeaways

Measure AI agent performance with objective metrics like success rate and latency.
Use both automated tests and human evaluation for subjective tasks.
Leverage asynchronous calls and streaming for efficient performance measurement.

Verified 2026-04 · gpt-4o, claude-3-5-sonnet-20241022

Verify ↗