Chapter 06
Week 2 Day 1
Notebook.venv39 cells
Week 2 Day 1
And now! Our first look at OpenAI Agents SDK
You won't believe how lightweight this is..
Three Parts to this lab
Part 1: A simple "Agent" and "Agent Loop"
Basically an LLM call. We'll add tracing and streaming to the mix.
Part 2: Adding a Tool
A familiar one, but oh-so-easy
Part 3: Adding Memory
So that different Agent calls know about each other
In [ ]python · cell 4
python
# The imports
import os
import requests
from dotenv import load_dotenv
from openai.types.responses import ResponseTextDeltaEvent
from agents import Agent, Runner, trace, function_tool, SQLiteSession
load_dotenv(override=True)Sidenote
The actual name of this framework on the official Python index pypi.org is openai-agents
So for your own projects in the future, you would do:
pip install openai-agents
or
uv add openai-agents
followed by
from agents import Agent, Runner, trace
Beware that doing a pip install agents would install something completely different - an older reinforcement learning library.
In [ ]python · cell 6
python
# Make an agent with name, instructions, model
agent = Agent(name="Jokester", instructions="You are a joke teller", model="gpt-5.4-mini")In [ ]python · cell 7
python
# Run the joke with Runner.run(agent, prompt)
result = await Runner.run(agent, "Tell a joke about Autonomous AI Agents")In [ ]python · cell 8
python
# Here is the final output
print(result.final_output)In [ ]python · cell 9
python
# Here is the detail of the LLM calls
result.to_input_list()Adding Observability with a trace
In [ ]python · cell 11
python
with trace("Telling a joke"):
result = await Runner.run(agent, "Tell a joke about Autonomous AI Agents")
print(result.final_output)Now go and look at the trace
In [ ]python · cell 13
python
# Streaming
result = Runner.run_streamed(agent, input="Please tell me 5 jokes about AI Agents.")
async for event in result.stream_events():
if event.type == "raw_response_event" and isinstance(event.data, ResponseTextDeltaEvent):
print(event.data.delta, end="", flush=True)Part 2: Adding a tool
In [ ]python · cell 15
python
pushover_user = os.getenv("PUSHOVER_USER")
pushover_token = os.getenv("PUSHOVER_TOKEN")
pushover_url = "https://api.pushover.net/1/messages.json"
if pushover_user:
if pushover_user.startswith("u"):
print("Pushover user found and looks good")
else:
print("Pushover user found but doesn't start with u")
else:
print("Pushover user not found")
if pushover_token:
if pushover_token.startswith("a"):
print("Pushover token found and looks good")
else:
print("Pushover token found but doesn't start with a")
else:
print("Pushover token not found")In [ ]python · cell 16
python
# Remember this?
def push(message):
print(f"Push: {message}")
payload = {"user": pushover_user, "token": pushover_token, "message": message}
requests.post(pushover_url, data=payload)In [ ]python · cell 17
python
push("HEY!!")In [ ]python · cell 18
python
pushIn [ ]python · cell 19
python
# Now this:
@function_tool
def push_tool(message: str) -> str:
""" Send the given message to the user as a push notification """
payload = {"user": pushover_user, "token": pushover_token, "message": message}
result = requests.post(pushover_url, data=payload).status_code
return f"Push sent with API status code {result}"In [ ]python · cell 20
python
push_toolIn [ ]python · cell 21
python
push_tool.descriptionIn [ ]python · cell 22
python
notifier = Agent(name="Notifier", model="gpt-5.4-mini", instructions="You notify the user upon request", tools=[push_tool])In [ ]python · cell 23
python
with trace("Pizza has arrived"):
result = await Runner.run(notifier, "Notify the user that the pizza is here")
print(result.final_output)Now go and look at the trace
Part 3: Sessions (memory)
Within a Runner.run() application level turn, the conversation history is maintained.
But each call to Runner.run() is a fresh start.
Let's see that:
In [ ]python · cell 26
python
agent = Agent(name="Assistant", model="gpt-5.4-mini")In [ ]python · cell 27
python
response = await Runner.run(agent, "Hi there. My name is Ed.")
print(response.final_output)In [ ]python · cell 28
python
response = await Runner.run(agent, "What's my name?")
print(response.final_output)Memory approach 1 - just manually pass in the list of dicts
In [ ]python · cell 30
python
response = await Runner.run(agent, "Hi there. My name is Ed.")
print(response.final_output)In [ ]python · cell 31
python
response.to_input_list()In [ ]python · cell 32
python
next_input = response.to_input_list() + [{"role": "user", "content": "What's my name?"}]
next_inputIn [ ]python · cell 33
python
response = await Runner.run(agent, next_input)
print(response.final_output)Another approach - use OpenAI Agents SDK built in SQLLite session
In [ ]python · cell 35
python
# This is created in-memory
# For an on-disk memory, use SQLiteSession("12345", "memory.db")
session = SQLiteSession("12346")In [ ]python · cell 36
python
response = await Runner.run(agent, "Hi there. My name is Ed.", session=session)
print(response.final_output)In [ ]python · cell 37
python
response = await Runner.run(agent, "What's my name?", session=session)
print(response.final_output)WOW
Can you believe how much we got done in Lab 1?!
Agents, Runner (Agent Loop), traces (Observability), Streaming, Function Tools, Memory!
Remember to check out the docs:
https://openai.github.io/openai-agents-python/
Even better news: many of the lightweight Agent Frameworks are very similar, so you practically know them all..