Python: [Bug]: WorkflowAgent.as_agent() concatenates all messages into one, losing message boundaries in handoff workflows

[frdeange]

Description

What happened?

When using workflow.as_agent() with a HandoffBuilder workflow, all messages from the conversation (including user input and agent responses) are concatenated into a single message with author_name='handoff-coordinator'.

The user's input text appears at the beginning of the assistant's response, making it impossible to distinguish between user messages and agent responses.

Example output:

Input: "I want to request sick leave"
Output: "I want to request sick leaveOf course! I can help you with that..."

What did you expect to happen?

The AgentRunResponse.messages should contain separate ChatMessage objects preserving:

• Individual role (user vs assistant)
• Individual author_name (which agent responded)
• Separate text content for each message

This works correctly when using workflow.run().get_outputs() directly, but NOT when using workflow.as_agent().run().

Steps to reproduce the issue

1. Create a handoff workflow with a coordinator and specialist agents
2. Call workflow.as_agent() to wrap it as an agent
3. Call agent.run(messages=[...])
4. Observe that response.messages contains a single message with all content concatenated

---

Code Sample

import asyncio
from agent_framework import ChatMessage, HandoffBuilder, ChatAgent
from agent_framework.azure import AzureOpenAIChatClient

async def reproduce_bug():
    client = AzureOpenAIChatClient(
        endpoint="YOUR_ENDPOINT",
        deployment_name="YOUR_MODEL", 
        api_key="YOUR_KEY"
    )
    
    coordinator = ChatAgent(
        chat_client=client,
        name="Coordinator",
        system_prompt="Route to SpecialistAgent for any request."
    )
    
    specialist = ChatAgent(
        chat_client=client,
        name="SpecialistAgent", 
        system_prompt="You help with requests. Respond briefly."
    )
    
    workflow = (
        HandoffBuilder(name="Test", participants=[coordinator, specialist])
        .set_coordinator(coordinator)
        .add_handoff(coordinator, [specialist])
        .with_interaction_mode(interaction_mode="autonomous", autonomous_turn_limit=2)
        .build()
    )
    
    # BUG: Using as_agent() concatenates everything
    agent = workflow.as_agent(name="TestAgent")
    response = await agent.run(messages=[ChatMessage(role="user", text="Hello")])
    
    print(f"Number of messages: {len(response.messages)}")
    for msg in response.messages:
        print(f"Author: {msg.author_name}, Role: {msg.role}")
        print(f"Text: {msg.text[:100]}...")  # Shows "Hello" + agent response concatenated
    
    # WORKS CORRECTLY: Using workflow.run() directly
    result = await workflow.run(message=[ChatMessage(role="user", text="Hello")])
    outputs = result.get_outputs()
    print(f"\nDirect workflow outputs: {len(outputs)} separate items")

asyncio.run(reproduce_bug())

Error Messages / Stack Traces

No error/exception is raised. The bug is in the behavior:

Using as_agent().run():
- response.messages = [ChatMessage(author='handoff-coordinator', text='HelloHi! How can I help?')]
  ^ User input "Hello" concatenated with assistant response

Using workflow.run().get_outputs():  
- outputs = [ChatMessage(role='user', text='Hello'), ChatMessage(role='assistant', author='SpecialistAgent', text='Hi! How can I help?')]
  ^ Correctly separated messages

Package Versions

agent-framework: 1.0.0b260107

Python Version

Python 3.12

Additional Context

Root Cause Analysis

The bug is in WorkflowAgent._convert_workflow_event_to_agent_update() in _agent.py (around line 315-350).

When HandoffCoordinator emits yield_output(list[ChatMessage]), the method:

1. Does NOT recognize list[ChatMessage] as a special case
2. Falls through to _extract_contents(data) which flattens all contents from all messages into a single list
3. Creates ONE AgentRunResponseUpdate with all contents combined and author_name='handoff-coordinator'

Current problematic code:

case WorkflowOutputEvent(data=data, source_executor_id=source_executor_id):
    if isinstance(data, ChatMessage):  # Only handles single ChatMessage
        return AgentRunResponseUpdate(...)
    
    # When data is list[ChatMessage], it falls here:
    contents = self._extract_contents(data)  # Flattens everything!
    return AgentRunResponseUpdate(
        contents=contents,  # All contents in ONE update
        author_name=source_executor_id,  # 'handoff-coordinator'
        ...
    )

Suggested fix:
Add handling for list[ChatMessage] before the fallback:

case WorkflowOutputEvent(data=data, source_executor_id=source_executor_id):
    # ... existing checks ...
    
    if isinstance(data, ChatMessage):
        return AgentRunResponseUpdate(...)
    
    # NEW: Handle list of ChatMessage - yield multiple updates
    if isinstance(data, list) and all(isinstance(m, ChatMessage) for m in data):
        # Return updates for each message separately, or handle differently
        # This needs design decision on how to yield multiple updates
        pass
    
    contents = self._extract_contents(data)  # Fallback

Workaround

Use workflow.run() directly instead of workflow.as_agent().run() and manage threads manually. This preserves message boundaries but loses the convenient thread/memory support of WorkflowAgent.