Structured Output + Tool Calling is not working with Gemini in Open AI Agents SDK #2257
Description
I'm trying this Agent with tools and structured output in Open AI Agents SDK with Gemini.
import os
import random
from typing import List, Literal
from agents import Agent, ItemHelpers, Runner, function_tool, set_tracing_disabled
from agents.extensions.models.litellm_model import LitellmModel
from pydantic import BaseModel, Field
set_tracing_disabled(disabled=True)
os.environ["GOOGLE_API_KEY"] = "..."
model = LitellmModel(model="gemini/gemini-3-flash-preview", api_key=os.getenv("GOOGLE_API_KEY"))
class Joke(BaseModel):
setup: str = Field(description="Joke setup")
punchline: str = Field(description="Joke punchline")
class JokesResponse(BaseModel):
count: int = Field(description="Total number of jokes generated")
style: Literal["dad", "dark", "absurd", "clean"] = Field(description="Chosen style")
jokes: List[Joke] = Field(description="List of jokes")
class JokePlan(BaseModel):
count: int = Field(ge=1, le=5, description="How many jokes to tell (1..5)")
style: Literal["dad", "dark", "absurd", "clean"] = Field(description="Joke style")
class JokerAgentToolStructured:
def __init__(self):
self.agent = Agent(
name="Joker+Structured",
instructions=(
"First, ALWAYS call the `make_joke_plan` tool.\n"
"Then generate the FINAL response as JSON strictly compatible with `JokesResponse`:\n"
"- `count` must match the plan\n"
"- `style` must match the plan\n"
"- `jokes` must contain exactly `count` elements\n"
"Keep the jokes in Spanish."
),
tools=[self.make_joke_plan],
output_type=JokesResponse,
model=model,
)
@staticmethod
@function_tool
def make_joke_plan() -> JokePlan:
"""Decide how many jokes and which style to use."""
return JokePlan(
count=random.randint(1, 5),
style=random.choice(["dad", "absurd", "clean", "dark"]),
)
async def run(self, input: str):
result = Runner.run_streamed(self.agent, input=input)
print("=== Run starting ===")
async for event in result.stream_events():
if event.type == "raw_response_event":
continue
elif event.type == "agent_updated_stream_event":
print(f"Agent updated: {event.new_agent.name}")
continue
elif event.type == "run_item_stream_event":
item = event.item
if item.type == "tool_call_item":
print("-- Tool was called")
elif item.type == "tool_call_output_item":
print(f"-- Tool output: {item.output}")
elif item.type == "message_output_item":
print(f"-- Message output:\n{ItemHelpers.text_message_output(item)}")
print("=== Run complete ===")
final: JokesResponse = result.final_output
print("\n=== Final structured output (Pydantic object) ===")
print(final.model_dump_json(indent=2))
async def main():
agent = JokerAgentToolStructured()
await agent.run("I want jokes for a work meeting.")
await main()Relevant log output
File c:\Users\sejohermo\AppData\Local\Programs\Python\Python312\Lib\site-packages\agents\models\chatcmpl_stream_handler.py:91, in ChatCmplStreamHandler.handle_stream(cls, response, stream)
89 state = StreamingState()
90 sequence_number = SequenceNumber()
---> 91 async for chunk in stream:
92 if not state.started:
93 state.started = True
File c:\Users\sejohermo\AppData\Local\Programs\Python\Python312\Lib\site-packages\litellm\litellm_core_utils\streaming_handler.py:2013, in CustomStreamWrapper.__anext__(self)
2010 except Exception as e:
2011 from litellm.exceptions import MidStreamFallbackError
-> 2013 raise MidStreamFallbackError(
2014 message=str(e),
2015 model=self.model,
2016 llm_provider=self.custom_llm_provider or "anthropic",
2017 original_exception=e,
2018 generated_content=self.response_uptil_now,
2019 is_pre_first_chunk=not self.sent_first_chunk,
2020 )
MidStreamFallbackError: litellm.ServiceUnavailableError: litellm.MidStreamFallbackError: litellm.BadRequestError: Vertex_ai_betaException BadRequestError - b'{\n "error": {\n "code": 400,\n "message": "Request contains an invalid argument.",\n "status": "INVALID_ARGUMENT"\n }\n}\n' Original exception: BadRequestError: litellm.BadRequestError: Vertex_ai_betaException BadRequestError - b'{\n "error": {\n "code": 400,\n "message": "Request contains an invalid argument.",\n "status": "INVALID_ARGUMENT"\n }\n}\n'