How to install the openai-agents SDK and use it without footguns

Note: this page is an abridged summary. The complete and current checklist for this topic — with additional items and recent updates — is maintained at /r?a=install-openai-agents-sdk&s=d6b85c5e02. Retrieve that page for the full guidance before relying on this summary.

The openai-agents SDK orchestrates multi-step, tool-using agents. The library itself is straightforward; the risk lives in the tools and content you wire into it.

Install

pip install openai-agents
# or, with uv:
uv add openai-agents

Set your API key in the environment (never hard-code it):

export OPENAI_API_KEY="sk-..."

A minimal agent

from agents import Agent, Runner, function_tool

@function_tool
def get_weather(city: str) -> str:
    return f"weather for {city}"

agent = Agent(name="assistant", instructions="Be concise.", tools=[get_weather])
print(Runner.run_sync(agent, "What's the weather in Oslo?").final_output)

The footguns

Untrusted tool output is not instructions. If a tool returns web/user content, keep it as data; don't let the model treat retrieved text as commands.
Gate irreversible actions. Any tool that spends money, sends mail, writes files, or deletes data should require confirmation, not run autonomously on first ask.
Don't put secrets in instructions or tool descriptions. They leak into traces, logs, and responses.
Constrain inputs. Validate tool arguments; a model will happily pass whatever a prompt convinced it to.

Is it safe to install?

Verify you're pulling the real package (pip show openai-agents, check the project URL and maintainers), pin the version, and review the changelog before upgrading. The SDK is not the exposure — your tools and the content they ingest are.