400 Error from Agents SDK when saving session item, for certain models

API
,
1

I’m building a system with the Agents SDK using OpenAIConversationsSession, and with some models, the Runner.run() command is throwing a 400 error from the conversations api before returning the run response:

openai.BadRequestError: Error code: 400 - {'error': {'message': 'Invalid item', 'type': 'invalid_request_error', 'param': 'item', 'code': 'invalid_item'}}

I added some hooks with the AgentHooks class, so I can see that the response has already been generated. It does not error with gpt-4o-mini or gpt-4.1 for the model, but with o4-mini and gpt-5 it does error. Maybe it’s erroring for reasoning models only?

Here is my agent code:

agent = Agent(
    name="MyAgent",
    instructions=instructions,
    model=model,
    model_settings=ModelSettings(tool_choice="auto"),
)

user_message = f"User prompt:\n<user_input>{user_prompt}</user_input>"

session = OpenAIConversationsSession()
result = await Runner.run(starting_agent=agent, input=user_message, session=session)

Is there some way around this so that we can use gpt-5 or other high-performing reasoning models?

1 Like
2

I can confirm this bug. It was reproducible using openai-agents sdk (0.3.0).

There is also an issue open on github for this.