Support for Gemini Flash

[bhavik-dand]

Are the latest Gemini models supported out of box?

[rm-openai]

Should work via OpenAI compatibility. See #110 for examples.

[dominicdev]

it works but only if no tools or handsoff, but the output is not using the tool

"The available tools lack the necessary functionality to get weather information. I need more information or different tools to answer your question."

`
from agents import Agent, Runner, AsyncOpenAI, OpenAIChatCompletionsModel,ModelSettings, function_tool

@function_tool
def get_weather(city: str) -> str:
print("call function")
return f"The weather in {city} is sunny"

gemini_agent = Agent(
name="Gemini agent",
instructions="You are helpful assistant",
tools=[get_weather],
model=OpenAIChatCompletionsModel(
model="google/gemini-1.5-flash-002",
openai_client=external_client,
),

model="gpt-3.5-turbo",

)

result = await Runner.run(gemini_agent, input="what is the weather in tokyo?")
print(result.final_output)

`

[mocheng]

It looks not working with tool. It throws 400 error.

@function_tool
def get_weather(city):
    """
    Get weather from city

    Args:
        city (str): city
    """
    print('### in get_weather city:', city)
    return f'Weather in {city} is sunny'

AGENT_INSTRUCTION = """ \
Always respond in haiku form
"""

agent = Agent(
    name='Assistant',
    instructions=AGENT_INSTRUCTION,
    tools=[get_weather],
    model=OpenAIChatCompletionsModel(
        model='gemini-2.0-flash',
        openai_client=external_client,
    )
)

query = "what's the weather in Boston?"
result = Runner.run_sync(agent, query)
print(result.final_output)

The error is

openai.BadRequestError: Error code: 400 - [{'error': {'code': 400, 'message': 'Request contains an invalid argument.', 'status': 'INVALID_ARGUMENT'}}]

If I remove the tools argument to Agent, then it has valid output like this:

I cannot say now.
I have no weather data
For fair Boston town.

[rm-openai]

@mocheng @dominicdev do you know what specifically about the tools is wrong? I'm happy to fix, but not super familiar with the gemini openai integration.

[rm-openai]

I tested on anthropic/together ai and tools worked fine

[hburrichter]

Maybe related to this problem: BerriAI/litellm#8351

Gemini doesn't like "default": null, within the tools JSON schema, which is added by Pydantic if the argument/attribute is nullable.

[rm-openai]

I just tested this with gemini and it worked fine: https://github.com/openai/openai-agents-python/blob/main/examples/model_providers/custom_example_global.py

[mocheng]

@rm-openai I figured out what went wrong. In my previous code snippet, the tool missed typing.

@function_tool
def get_weather(city):

It works after changing to be like this:

@function_tool
def get_weather(city: str):

[github-actions]

This issue is stale because it has been open for 7 days with no activity.

[jaijuneja]

I just tested this with gemini and it worked fine: https://github.com/openai/openai-agents-python/blob/main/examples/model_providers/custom_example_global.py

I'm having a bit of an odd experience with this on my end. If I run the code 10 times it will work maybe 2-3 times out of 10 (i.e. it calls the tool and returns a response), and the rest of the time it doesn't call the tool and hallucinates. Have tried with flash-2.0 and 2.5-pro.

I also find that if I have multiple tools available it almost never calls more than one tool, even if there is strong logical basis to call multiple tools.

I looked this up and it appears to be a common complaint with Gemini. One of the suggestions to improve this is to set the config var "function_calling_config": {"mode": "any"}.

https://ai.google.dev/gemini-api/docs/function-calling?example=meeting#single-turn-any-mode

I wonder if there's any way to enforce that configuration directly from the Agents SDK, or whether it just uses the defaults?

[rm-openai]

@jaijuneja From the gemini docs, it seems like "function_calling_config": {"mode": "any"}. means:

ANY: The model is constrained to always predict a function call and guarantee function schema adherence.

The equivalent in the Agents SDK is to set ModelSettings(tool_choice="required")