OpenAI Agents SDK

pip install openai-agents
export API_KEY="sk-at-…"

An agent with a tool#

import os

from agents import Agent, ModelSettings, OpenAIChatCompletionsModel, Runner, function_tool, set_tracing_disabled
from openai import AsyncOpenAI

# Traces go to OpenAI's servers by default: switch them off (see Tracing below).
set_tracing_disabled(True)

client = AsyncOpenAI(
    base_url="https://api.aitokens.ch/v1",
    api_key=os.environ["API_KEY"],
)


@function_tool
def get_weather(city: str) -> str:
    """Current weather for a city."""
    return f"{city}: 14 °C, light rain"


agent = Agent(
    name="Assistant",
    instructions="Answer in one sentence.",
    model=OpenAIChatCompletionsModel(model="qwen3.8-27b", openai_client=client),
    model_settings=ModelSettings(max_tokens=2048),
    tools=[get_weather],
)

result = Runner.run_sync(agent, "What is the weather in Lugano?")
print(result.final_output)

What came back on qwen3.8-27b:

The weather in Lugano is currently 14 °C with light rain.

The agent called get_weather, read its result and answered with it.

Chat completions for every agent#

The SDK uses OpenAI's Responses API by default. Give each agent an OpenAIChatCompletionsModel, as above, or set it once for every agent:

from agents import set_default_openai_api, set_default_openai_client

set_default_openai_client(client, use_for_tracing=False)
set_default_openai_api("chat_completions")

We ran the same agent through the Responses API too, on qwen3.6-35b. It answered "I do not have access to real-time weather data": AI Tokens's Responses API did not read the tool's result on 11 October 2026, and it has no streaming.

Tracing#

Unless tracing is disabled, the SDK uploads traces to OpenAI's servers. It signs them with the key in OPENAI_API_KEY, or with the key of the client given to set_default_openai_client. Either can be your AI Tokens key, which would then be sent to OpenAI. Call set_tracing_disabled(True), or pass use_for_tracing=False to set_default_openai_client and send traces to a processor of your own.

Limits#

• max_tokens up to 8192. A higher value is refused with 422.
• Models with tools: qwen3.8-27b, qwen3.6-35b, qwen3-coder-30b, gpt-oss-120b, gemma-4-26b and ministral-3-14b returned tool calls on 11 October 2026. See Which model.