microsoft
Multiple tool call failure with Gemini
What happened?
Describe the bug
Use of gemini models with multiple tool calls leads to openai.BadRequestError. This happens only if I set model_client_stream=True
To Reproduce
Following snippet will reproduce the issue.
from autogen_core.tools import BaseTool, FunctionTool
from autogen_agentchat.agents import AssistantAgent
from autogen_agentchat.messages import TextMessage
from typing import List
from autogen_ext.models.openai import OpenAIChatCompletionClient
from autogen_core import CancellationToken
import asyncio
async def get_current_time() -> str:
return "The current time is 12:00 PM."
async def get_stock_price(stock: str ="MFST") -> str:
return "The stock price is 500 dollors."
async def main():
tools: List[BaseTool] = [
FunctionTool(
get_current_time,
name="get_current_time",
description="Get current time in Paris.",
),
FunctionTool(
get_stock_price,
name="get_stock_price",
description="Fetch the stock price for a given company."
),
]
model_client = OpenAIChatCompletionClient(
model="gemini-2.5-flash",
api_key="",
)
response = await AssistantAgent(
name="Agent",
model_client=model_client,
model_client_stream=True,
reflect_on_tool_use=True,
tools=tools,
max_tool_iterations=10,
system_message= "You are an assistant with some tools that can be used to answer some questions",
).on_messages([TextMessage(content="What is current time of Paris now? And what is stock price of MSFT", source="user")], CancellationToken())
print(response)
if __name__ == "__main__":
asyncio.run(main())
Which packages was the bug in?
Python AgentChat (autogen-agentchat>=0.4.0)
AutoGen library version.
Python 0.6.4
Other library version.
No response
Model used
gemini-2.5-flash
Model provider
Google Gemini
Other model provider
No response
Python version
3.11
.NET version
None
Operating system
Ubuntu
It seems that the Gemini server generates malformed tool call argument that looks like this {}{"stock":"MSFT"}. The first chunk comes in is {}, and then the second chunk is {"stock":"MSFT"}.
I have the same problem VLLM - QWEN. model_client_stream=True, reflect_on_tool_use=True,
as soon as model_client_stream is activated, HandOff no longer works without errors.
autogen-agentchat 0.6.4 arguments=“” is always an empty string.
The HandOff really only happens after Streem has been terminated.
model_client_stream=False So everything works perfectly.
Especially if you want to ask the user a question, this is very annoying. I hope there is a possibility to fix the problem.
was getting same error before. so i have used pydantic model as input not just parameters. till now didn't got multiple tool call based error.
class IPParms(Basemodel):
stock: str = Field()
def stocl_tool(params: IPParams):
pass
Edit: got error today even after using pydantic models
