Invalid finish_reason error in tool calls with anthropic endpoint

[ravishqureshi]

What happened?

Models such as anthropic support only sequential function calling. Thus when we pass multiple tools while using an anthropic model, the function caling of subsequent tools is not handled.

and if we try to handle it ourselves, it fails again at finish_reason
ValidationError: 1 validation error for CreateResult
finish_reason
Input should be 'stop', 'length', 'function_calls', 'content_filter' or 'unknown' [type=literal_error, input_value='end_turn', input_type=str]
For further information visit https://errors.pydantic.dev/2.10/v/literal_error

What did you expect to happen?

Either autogen allows the developer to handle the sequential call. Or autogen handles sequential call itself the way it handles parallel tool calls for openai

How can we reproduce it (as minimally and precisely as possible)?

async def run_task(agent, messages):
"""
Sends messages to the agent and handles tool calls.

Args:
    agent (AssistantAgent): The initialized AssistantAgent.
    messages (list): The conversation history.

Returns:
    str: The final response after tool execution.
"""
# Send messages to the agent
cancellation_token = CancellationToken()

response = await agent.on_messages(messages=messages,cancellation_token=cancellation_token)
print("here 0")
print(response)
if response.inner_messages:
    print("here 0.1")
    for message in response.inner_messages:
        print(response.chat_message.content)
        print(type(response.chat_message.content))
        print("*********")
        if isinstance(message, ToolCallRequestEvent):
          tool_result_message = ToolCallSummaryMessage(
                      content=str(response.chat_message.content),
                      source="tool"
                  )
          response = await agent.on_messages(messages=[tool_result_message], cancellation_token=cancellation_token)
          break
print("here 4")
return response.chat_message.content

Above code works for openai. for anthropic, i will need to remove the break and handle subsequent function calling which results in failure

AutoGen version

0.4.2

Which package was this bug in

AgentChat

Model used

sonnet3.5

Python version

3.12

Operating system

No response

Any additional info you think would be helpful for fixing this bug

No response

[rysweet]

Thanks for the report @ravishqureshi - we will have a look and think about the right approach. Feel free to also consider submitting a pr.

[ekzhu]

@ravishqureshi the AssistantAgent is designed to keep each turn limited to 1 model client call, and 2 if reflect_on_tool_use=True. Consider the following, if you want to execute multiple tools.

1. Call the same agent again if the response.chat_message is a ToolResultSummaryMessage. This could be an orchestration decision by the team, e.g., using custom selector function in SelectorGroupChat, or just use Swarm team.
2. Breaking up into multiple agents and spread the tools.

[ekzhu]

PR #5085 to fix the finish_reason bug

[ekzhu]

@ravishqureshi if you want to disable parallel tool calls all together, set parallel_tool_calls=False in the OpenAIChatCompletionClient.

[ravishqureshi]

Thanks for a quick turnaround on the fix! I really love how the issues are fixed at speed by MS. Keep up the good work.

Also, i wanted to actually allow "parallel_tool_calls" for models like claude that do only sequential tool call for now. But this fix above will allow me to handle these cases easily..

[ekzhu]

i wanted to actually allow "parallel_tool_calls" for models like claude that do only sequential tool call for now

Claude can support parallel tools: https://docs.anthropic.com/en/docs/build-with-claude/tool-use#multiple-tool-example

I am not familiar with it, but it seems to me that you may be able to get "consistent" agent behavior -- note that model is different -- from both Claude and OpenAI models.

Keep up the good work.

Thank you!