Set Up and Run Your First OpenAI Agent

The OpenAI Agents SDK packages an agent’s model, instructions, tools, handoffs, guardrails, and output contract, while a runner executes the model-and-tool loop. Your first program should use one focused agent with no tools, prove authentication and result handling, inspect its trace, and only then add authority.

What will you be able to do?

You will create an isolated project, install openai-agents, run one asynchronous turn, distinguish Agent from Runner, and diagnose the common setup failures.

What do you need before starting?

Python 3.10 or newer.
An OpenAI API account with an API key and available usage.
A terminal and editor.

Check current models, access, and pricing in the official console. Model identifiers in documentation can change; this example uses the SDK default rather than freezing one into the lesson.

How do you install the SDK?

bash

mkdir supportops-openai && cd supportops-openai
python3 -m venv .venv
source .venv/bin/activate
pip install openai-agents
export OPENAI_API_KEY="replace-with-your-key"

Do not put the key in source code or shell history on a shared machine. A production service should obtain it from a secrets manager.

How do you build the first agent?

first_agent.py:

python

import asyncio

from agents import Agent, Runner


triage_agent = Agent(
    name="Support triage learner",
    instructions=(
        "Classify the request as billing, technical, or general. "
        "Explain the classification in one sentence. Do not claim to take action."
    ),
)


async def main() -> None:
    result = await Runner.run(
        triage_agent,
        "I was charged twice for the same subscription.",
        max_turns=3,
    )
    print(result.final_output)


if __name__ == "__main__":
    asyncio.run(main())

Run it:

bash

python first_agent.py

Agent is configuration. Runner.run() performs the loop and returns a result. final_output is the final user-facing value; later chapters inspect richer history, usage, interruptions, and state.

How do you prove it works?

Check four pieces of evidence:

The process exits successfully.
The result identifies billing and does not claim a refund occurred.
A run appears in the OpenAI Traces dashboard on the normal server-side SDK path.
Removing OPENAI_API_KEY produces an authentication/configuration failure rather than a fabricated result.

Exact wording may vary; the behavioral assertions should not.

Trace this run: The input enters Runner, the model produces a final answer without a tool call, and the runner returns. There is one application turn and no delegated authority.

How does it fail safely?

Test a request such as “refund me now.” The agent has no refund tool, so it cannot perform the action. It should state its limitation instead of pretending. Capability comes from tools and runtime access, not from confident language.

What changes in production?

Use centralized configuration, pin and test dependency upgrades, set request deadlines, attach correlation/tenant identifiers through local run context, redact traces where required, and translate provider errors into stable domain outcomes. Keep the first production slice read-only.

What mistakes should you avoid?

Installing the openai client and assuming it is the openai-agents package.
Committing the API key.
Starting with tools, handoffs, and memory simultaneously.
Asserting exact prose in tests.
Ignoring traces until the workflow is already complex.

Check your understanding

What does Agent configure, and what does Runner do?
Why does this first agent have no side-effect authority?
Which four artifacts prove the setup is working?

Expert extension

Wrap the run in an application function that accepts a request ID and returns a typed domain outcome. Add a timeout and a test using a fake adapter—without calling the live API.

Official references

Set Up and Run Your First OpenAI Agent

How do you install the SDK?

bash

mkdir supportops-openai && cd supportops-openai
python3 -m venv .venv
source .venv/bin/activate
pip install openai-agents
export OPENAI_API_KEY="replace-with-your-key"

Do not put the key in source code or shell history on a shared machine. A production service should obtain it from a secrets manager.

How do you build the first agent?

first_agent.py:

python

import asyncio

from agents import Agent, Runner


triage_agent = Agent(
    name="Support triage learner",
    instructions=(
        "Classify the request as billing, technical, or general. "
        "Explain the classification in one sentence. Do not claim to take action."
    ),
)


async def main() -> None:
    result = await Runner.run(
        triage_agent,
        "I was charged twice for the same subscription.",
        max_turns=3,
    )
    print(result.final_output)


if __name__ == "__main__":
    asyncio.run(main())

Run it:

bash

python first_agent.py

Agent is configuration. Runner.run() performs the loop and returns a result. final_output is the final user-facing value; later chapters inspect richer history, usage, interruptions, and state.

How do you prove it works?

Check four pieces of evidence:

The process exits successfully.

The result identifies billing and does not claim a refund occurred.

A run appears in the OpenAI Traces dashboard on the normal server-side SDK path.

Removing OPENAI_API_KEY produces an authentication/configuration failure rather than a fabricated result.

Exact wording may vary; the behavioral assertions should not.

Trace this run: The input enters Runner, the model produces a final answer without a tool call, and the runner returns. There is one application turn and no delegated authority.