Building Your First AI Agent in Python with the OpenAI Agents SDK

A tutorial demonstrates building a Python research agent with the OpenAI Agents SDK, using the SDK's Runner to manage the tool-calling loop, a hosted WebSearchTool for live web lookups, and a custom @tool-decorated Python function for calculations. In the walkthrough's sample run, the agent reported Python 3.14.7 as the latest stable maintenance release and said Python 3.15's final release is scheduled for October 1, 2026. The setup requires installing the openai-agents package via uv or pip and setting an OpenAI API key.

Building Your First AI Agent in Python with the OpenAI Agents SDK Build your first AI agent in Python with the OpenAI Agents SDK, adding web search, Python tools, conversation memory, and tracing in a simple end-to-end workflow. Building an AI agent in Python is now much simpler than it used to be. With the OpenAI Agents SDK , you can create agents, connect tools, maintain conversation state, and run multi-step workflows without building your own tool-calling loop from scratch. In this tutorial, we will build a practical research agent that can search the web, call a custom Python function, remember previous messages, and decide when to use each capability . What Is an AI Agent? A standard large language model LLM application usually follows a simple flow: Prompt — Model — Response The model receives a prompt, generates an answer, and stops. An AI agent adds tools and an execution loop : User — Agent — Tool if needed — Agent — Final Response Instead of only generating text, the agent can decide whether it needs to take an action first, such as searching the web, calling a Python function, or retrieving information from memory. With the OpenAI Agents SDK, the Runner manages this process automatically. It sends the request to the model, executes any requested tools, passes the results back to the agent, and continues until the agent produces a final answer. This ability to reason, use tools, and continue working across multiple steps is what makes an agent different from a basic LLM application. Set Up the Project Before we build the agent, let's create a clean Python environment and install everything we need. Create a Python project with uv : mkdir first-ai-agent cd first-ai-agent uv init uv add openai-agents Or install the package with pip : pip install openai-agents Set your OpenAI API key: export OPENAI API KEY="your-api-key" On Windows, export works in Git Bash. In Command Prompt or PowerShell, you can use: setx OPENAI API KEY "your-api-key" Build Your First Agent Now we can turn a few lines of Python into an agent that decides when it needs fresh information from the web. Create main.py : python from agents import Agent, Runner, WebSearchTool agent = Agent name="Research Assistant", instructions= "You are a concise research assistant. " "Use web search when current information is required." , tools= WebSearchTool , result = Runner.run sync agent, "What are the latest developments in Python?" print result.final output Run it: uv run main.py Output: As of September 14, 2026 , the major Python developments are: - Python 3.14 is the latest stable feature series ; the newest maintenance release is 3.14.7 . It adds officially supported free-threaded builds, deferred annotation evaluation, template string literals t"" , ... - Python 3.15 is nearing release : the final release is scheduled for October 1, 2026 . ... For new projects, Python 3.14.7 is currently the safest default. WebSearchTool is a hosted OpenAI tool, so the model can decide when it needs current information instead of you manually calling a search API. For example, our run correctly identified Python 3.14.7 as the current stable release. Add Your Own Python Tool Web search is useful, but agents become much more powerful when they can also interact with your own Python code. Add a calculation tool: python from agents.decorators import tool @tool def calculate total price: float, quantity: int - float: """Calculate the total cost of an order.""" return price quantity Then give it to the agent: agent = Agent name="Assistant", instructions="Help with research and calculations.", tools= WebSearchTool , calculate total, , Test it: result = Runner.run sync agent, "I need 14 GPUs costing $2.49 each per hour. " "What will they cost per hour?" print result.final output Output: 14 GPUs x $2.49/hour = $34.86 per hour . The SDK automatically uses the function name, arguments, type hints, and docstring to create a tool schema the model can understand. You do not need to manually define JSON schemas or write the tool-calling loop yourself. Tip: Write clear docstrings because they help the model understand what a tool does and when it should use it. Adding Conversation Memory A useful agent should not forget everything the moment the next message arrives, so let's give it conversation memory. To make the agent conversational, add a session: python from agents import SQLiteSession session = SQLiteSession "user 1", "conversations.db" Then pass it into each run: while True: message = input "You: " if message.lower in {"exit", "quit"}: break result = Runner.run sync agent, message, session=session, print "Agent:", result.final output Now try: The second request automatically receives the earlier conversation, so the agent understands what "the latest version" refers to. Because we used a file-backed SQLite database, the history can also persist after restarting the program. For larger applications, the SDK also provides session backends for systems such as Redis, SQLAlchemy, and MongoDB . How the Complete Agent Works At this point, all the pieces come together: The agent decides what action to take , the Runner executes the agent loop and any requested tools , and the session preserves conversation history between requests. Here is the complete main.py : python import os import sys from agents import Agent, Runner, SQLiteSession, WebSearchTool from agents.decorators import tool @tool def calculate total price: float, quantity: int - float: """Calculate the total cost of an order.""" return price quantity def main - None: if not os.environ.get "OPENAI API KEY" : sys.exit "Error: set the OPENAI API KEY environment variable first." agent = Agent name="Assistant", instructions=""" Help the user with research and calculations. Use the available tools when necessary. """, tools= WebSearchTool , calculate total, , session = SQLiteSession "user 1", "conversations.db" try: while True: try: message = input "You: " except EOFError, KeyboardInterrupt : print break command = message.strip .lower if command in {"exit", "quit"}: break if not command: continue try: result = Runner.run sync agent, message, session=session, except Exception as exc: print f"Agent error " f" {type exc . name } : {exc}" continue print "Agent:", result.final output finally: session.close if name == " main ": main This gives us a small but complete conversational agent with web search, a custom Python tool, persistent memory, basic error handling, and a continuous chat interface . The Agents SDK also enables tracing by default , which is especially useful once your workflows become more complex. Traces let you inspect the model calls, tool usage, handoffs, and other steps taken during an agent run. You can inspect these traces in the OpenAI dashboard to understand how the agent reached its answer, which tools it selected, and where problems occurred. This makes tracing particularly useful for debugging tool-calling behavior and improving agent instructions. Final Thoughts Building your first AI agent is less about creating a complex autonomous system and more about giving a model the ability to take useful actions . In this guide, we started with a simple agent, added web search , connected a custom Python tool , introduced persistent conversation memory , and brought everything together in a complete application with tracing and error handling. That foundation is enough to build surprisingly capable workflows. From here, you can gradually add APIs, MCP servers, guardrails, handoffs, more specialized tools, and additional agents as your application grows. The pattern stays the same: give the agent the right tools, let it decide when to use them, and let the Agents SDK handle the execution loop . Start small, make one workflow useful, and expand only when you actually need more complexity. \ Abid Ali Awan\ https://abid.work https://abid.work @1abidaliawan https://www.linkedin.com/in/1abidaliawan is a certified data scientist professional who loves building machine learning models. Currently, he is focusing on content creation and writing technical blogs on machine learning and data science technologies. Abid holds a Master's degree in technology management and a bachelor's degree in telecommunication engineering. His vision is to build an AI product using a graph neural network for students struggling with mental illness.