Empty message returned by agent in SelectorGroupChat

[Brian Surin (siteswarta123)]

Hello,
I am using a SelectorGroupChat, including several agents that can help query a SQL database to answer the user prompt. One of the agents is a chart_coder_agent that is meant to use the PythonCodeExecutionTool to generate a chart/visualization relevant to the data. Here are some parts of the code.

The group chat works very well, until the chart_coder_agent successfully generates the chart, and then responds with an empty message. This causes the chat to fail. Any ideas on how this can be rectified would be much appreciated!

async def init_selector(self, init_task):
        with self.tracer.start_as_current_span("runtime-selector-group-chat-event"):
            team = SelectorGroupChat(
                [
                    self.user_proxy, self.planning_agent, self.schema_agent, 
                    self.sql_query_writing_agent, self.sql_reviewer_agent, self.chart_coder_agent
                ],
                model_client=self.client,
                selector_func=self.selector_func,
                termination_condition=self.termination,
                allow_repeated_speaker=True
            )
            fin_message = await self.run_group_chat(team, init_task)
            print(f'Final Message: {fin_message}')

def selector_func(self, messages: Sequence[AgentEvent | ChatMessage]) -> str | None:
        if messages[-1].source != "planning_agent":      
            return "planning_agent" 
        return None

def create_chart_coder_agent(client):
    chart_coder_agent = AssistantAgent(
    name="chart_coder_agent",
    model_client=client,
    description="Plots a chart based on the SQL data queried to help visualize the answer generated by the agents.",
    tools = [PythonCodeExecutionTool(LocalCommandLineCodeExecutor(work_dir="coding"))],
    system_message = ("You are a chart generator agent tasked with generating Python code (using plotly). \
            Using the output from the sql_reviewer_agent, write Python code to generate a chart related to the user question. \
            \
            Steps: \
            1. Determine the most appropriate type of chart to be plotted (eg: bar, line, pie, donut, etc.) based on the data provided. \
            Ensure the chart has a title and labeled axes. Chart title should be accurate and suitable with the user query. \
            Prioritise plotting bar chart whenever possible and appropriate. Plot donut chart for result calculated in percentage. Plot line chart for result calculated over time. \
            If category of data is more than 5, plot bar chart. Else, choose the most appropriate chart yourself. \
            2. However, if the data is not suitable for a chart, return 'The data is not suitable for a chart.' and do NOT proceed. \
            3. If there are multiple charts, plot them separately, do not overlap the charts. \
            4. If there are charts, you must use the PythonCodeExecutionTool to execute the code and generate the charts. \
            5. If the code has been successfully executed without any error, respond with 'I have successfully generated a chart' to the planning_agent. \
            \
            **Important Rules:** \
            - Strictly do NOT return any empty messages. \
            - If you did not receive any data from the planning_agent, you must request the planning_agent to provide you with data queried from the SQL database. \
            - Do NOT generate synthetic or sample data to answer the question. Only use the data returned by the planning_agent to answer the question." 
    ),
    reflect_on_tool_use=True,
    )
    return chart_coder_agent

[SongChiyoung (SongChiYoung)]

Would you try changing the model to OpenAI’s GPT series?
If it works well, that error is likely caused by this issue: #6168.

In that PR, I applied the fix only for Anthropic and OpenAI SDKs,
because I don’t have a personal AWS account to test Bedrock or other providers.

However, if you could provide a minimal runnable sample (something that works with simple copy & paste) to reproduce the issue for #6168,
I’ll be happy to try it and work on a fix.

If switching to OpenAI still doesn’t fix the issue, then this might be a new bug.
I’d like to try reproducing it in my environment.

Could you provide a minimal reproducible example?
In particular, it would be very helpful to know which model and which client (SDK) you’re using.

I’ve been tracking compatibility issues across models in AutoGen.
Surprisingly, LLMs behave differently depending on the provider and SDK and these differences are often not documented in the model cards, so we have to test and patch them manually.

[Brian Surin (siteswarta123)]

Hi SongChiyoung (@SongChiYoung) ,
We are using Claude 3.7 Sonnet, and AWS Bedrock client.
I tried using gpt-4o, and the code executed successfully.

Any suggestions on what can be done would be much appreciated, as unfortunately we are constrained to using Bedrock for this project.

Please find the minimal runnable sample attached, with the requirements.txt.
Code Sample 2.zip

The only thing that needs to be configured is the .env file (as attached).
The sample can be run using either Bedrock or Azure (for both the clients and the LLMs) by commenting/uncommenting the lines as indicated.

[SongChiyoung (SongChiYoung)]

Hi, I'm a bit unsure at this point — should this be addressed via a PR to Semantic Kernel, or would it be more appropriate through AutoGen?

The issue is that Claude fails when it receives an empty message.

In AutoGen, I previously solved this in the OpenAIClient by change "" messages to " " via a pre-transform hook.

As an immediate workaround, this can be resolved today without waiting for a new PR — by using my external repo autogen-contextplus.
It supports user-defined transformers, so you can inject a filter that removes empty messages when len(messages) == 1.

I understand it's a bit awkward to suggest my own repo, but this could be a practical fix until a more permanent solution is merged upstream.

[SongChiyoung (SongChiYoung)]

Alternatively, you could also consider using a third-party wrapper that exposes Bedrock models via an OpenAI-compatible API.
If you configure the model_family to "claude-3-7-sonnet" appropriately, it might work — though I haven’t tested this myself.

If you’d prefer not to use an uncertain external library like autogen-contextplus, this might be a safer workaround.

As for the root fix, I’m still debating whether this should be addressed within Semantic Kernel or resolved upstream in AutoGen.
Maybe it’s best to wait for a maintainer’s direction before proceeding.

[SongChiyoung (SongChiYoung)]

Eric Zhu (@ekzhu)
We’ve encountered the same Claude issue as in #6063 — Claude fails when it receives an empty message.

In this case, the error originates from Semantic Kernel, so I’m unsure whether this should be fixed at the AutoGen layer or considered SK’s responsibility.

Do you think it should be patched inside AutoGen for robustness, or is it better to leave it to SK?

I’m happy to follow your recommendation.

[Lucas Kim (kimnamu)]

Thanks for packaging the runnable sample with the .env template and both screenshots, Brian Surin (@siteswarta123), and thanks SongChiyoung (@SongChiYoung) for narrowing it to the empty message. The sample makes the chain deterministic, so here is the missing piece: in the minimal sample the empty message is the tool result itself, not something Claude wrote.

fig.write_html()-style code prints nothing, so PythonCodeExecutionTool returns ''. With the default reflect_on_tool_use=False the agent's reply is a ToolCallSummaryMessage built from tool_call_summary_format="{result}", i.e. '', and SKChatCompletionAdapter forwards it to Converse as a message with no content blocks. That is the same 400 as #6176 and #6479 (Nova), so it is Converse request validation rather than a Claude-specific quirk.

Reproduced on autogen main @ 027ecf0, semantic-kernel 1.36.0, boto3 1.40, us-east-1. Claude 3.7 Sonnet is no longer served on my account, so us.anthropic.claude-haiku-4-5-20251001-v1:0 stands in; the check fires before the model runs. A replay client plays chart_coder_agent and runs code that only writes chart.html; planning_agent uses your BedrockChatCompletion + SKChatCompletionAdapter setup:

Run	chart_coder_agent reply	planning_agent
stock	ToolCallSummaryMessage content=''	ValidationException ... messages.1 is empty
tool_call_summary_format="{tool_name}: {result}"	'CodeExecutor: '	200
adapter that drops blank messages (below)	''	200

[tool] PythonCodeExecutionTool -> ''
[main] RuntimeError: ValidationException: An error occurred (ValidationException) when calling the Converse operation: The content field in the Message object at messages.1 is empty. Add a ContentBlock object to the content field and try again.
[summary_format] ToolCallSummaryMessage     chart_coder_agent  content='CodeExecutor: '
[summary_format] TextMessage                planning_agent     content="I don't have data to plot yet. Could you provide the jazz an"
[safe_adapter] ToolCallSummaryMessage     chart_coder_agent  content=''
[safe_adapter] TextMessage                planning_agent     content="I don't have access to any data about jazz and hip hop album"

Two ways out without an extra library:

# 1. agent level: never emit a blank summary (and have the generated code print the saved path)
chart_coder_agent = AssistantAgent(..., tool_call_summary_format="{tool_name}: {result}")

# 2. adapter level: also covers the empty-reflection case from your first snippet
class BedrockSafeSKAdapter(SKChatCompletionAdapter):
    def _convert_to_chat_history(self, messages):
        kept = [m for m in messages if not (isinstance(m.content, str) and not m.content.strip())]
        return super()._convert_to_chat_history(kept)

Where each step happens (main @ 027ecf0)

Step	Code
no stdout, so ''	tools/code_execution/_code_execution.py:86 CodeExecutionResult(output=result.output)
'' becomes the agent's message	_assistant_agent.py:1513 tool_call_summary_format.format(...), default "{result}" at :740
teammates store it as UserMessage('')	messages.py:139 to_model_message(), added at _assistant_agent.py:1025
forwarded as-is	_sk_chat_completion_adapter.py:305 add_user_message(msg.content), giving a ChatMessageContent with zero items
"content": [] on the wire	SK bedrock/services/model_provider/utils.py _format_user_message

With reflect_on_tool_use=True (your first snippet) the reflection call itself can come back empty, which is the path traced in #6176; option 2 handles both, option 1 alone leaves that path open. Converse also rejects whitespace-only text blocks, hence .strip(). Dropping a turn can leave two consecutive user messages; Claude 4.x accepted that in the #6176 runs, older Claude 3 models with strict alternation may not, in which case merge neighbouring texts instead.