How to get workflow outputs from workflow agent?

[cecheta]

I've been playing around with workflows. When running a workflow directly, it is easy to get the outputs, for example using .get_outputs(). But I don't know how to do it if you convert the workflow into an agent.

For example, using this workflow:

import asyncio

from agent_framework import (
    AgentExecutorRequest,
    ChatAgent,
    ChatMessage,
    WorkflowBuilder,
    WorkflowContext,
    executor,
)
from agent_framework.azure import AzureOpenAIChatClient
from azure.identity import AzureCliCredential


@executor
async def start(messages: list[ChatMessage], ctx: WorkflowContext[AgentExecutorRequest, str]) -> None:
    await ctx.yield_output("Start...")

    # Send the prompt to the next executor.
    await ctx.send_message(
        AgentExecutorRequest(
            messages=messages,
            should_respond=True,
        )
    )


def create_writer_agent() -> ChatAgent:
    """Factory function to create a writer agent."""
    chat_client = AzureOpenAIChatClient(credential=AzureCliCredential())
    return chat_client.create_agent(
        instructions=(
            "You are an excellent content writer. You create new content and edit contents based on the feedback."
        ),
        name="writer",
    )


def create_reviewer_agent() -> ChatAgent:
    """Factory function to create a reviewer agent."""
    chat_client = AzureOpenAIChatClient(credential=AzureCliCredential())
    return chat_client.create_agent(
        instructions=(
            "You are an excellent content reviewer."
            "Provide actionable feedback to the writer about the provided content."
            "Provide the feedback in the most concise manner possible."
        ),
        name="reviewer",
    )


async def main():
    # Build the workflow using the fluent builder.
    # Set the start node and connect an edge from writer to reviewer.
    workflow = (
        WorkflowBuilder()
        .register_executor(lambda: start, "start")
        .register_agent(create_writer_agent, "writer")
        .register_agent(create_reviewer_agent, "reviewer", output_response=True)
        .set_start_executor("start")
        .add_edge("start", "writer")
        .add_edge("writer", "reviewer")
        .build()
    )

I can get the outputs when running the workflow directly:

    prompt = "Create a slogan for a new electric SUV that is affordable and fun to drive."

    res = await workflow.run([ChatMessage(role="user", text=prompt)])
    for output in res.get_outputs():
        print(f"Workflow Output: {output}")

Workflow Output: Start...
Workflow Output: Clear and positive message; highlights affordability and fun. Consider shortening or using a more direct call-to-action (e.g., "Affordable Thrills. Electric Adventures."). Including the word "SUV" may improve clarity.

But how do you get the outputs if you run it as an agent?

    workflow_agent = workflow.as_agent()
    thread = workflow_agent.get_new_thread()

    res = await workflow_agent.run(prompt, thread=thread)

res.text looks like this, but it includes the output from writer, despite output_response not being set:

Start...Strong energy and appeal; nicely emphasizes both excitement and affordability. Consider including a nod to the SUV aspect or capability (e.g., "Charge Up the Fun—Adventure Meets Affordability," or "Electrify Your Everyday Adventure"). Overall, great balance of concise message and inviting tone.Charge Up the Fun—Adventure Within Reach.Strong energy and appeal; nicely emphasizes both excitement and affordability. Consider including a nod to the SUV aspect or capability (e.g., "Charge Up the Fun—Adventure Meets Affordability," or "Electrify Your Everyday Adventure"). Overall, great balance of concise message and inviting tone.

If you loop through res.messages, there are four messages, and again it includes the output from the writer:

Start...

Actionable Feedback:  
- Strong use of wordplay with "Charge Ahead."
- Clearly communicates affordability and fun, aligning with target messaging.
- To enhance SUV appeal, consider incorporating phrases that evoke spaciousness, freedom, or versatility, e.g., “Charge Ahead—Affordable Fun for Every Journey.”
- Test the slogan’s distinctiveness to avoid overlap with existing auto brands.

Charge Ahead—Affordable Adventure, Electrifying Fun.

Actionable Feedback:  
- Strong use of wordplay with "Charge Ahead."
- Clearly communicates affordability and fun, aligning with target messaging.
- To enhance SUV appeal, consider incorporating phrases that evoke spaciousness, freedom, or versatility, e.g., “Charge Ahead—Affordable Fun for Every Journey.”
- Test the slogan’s distinctiveness to avoid overlap with existing auto brands.

[moonbox3]

Hi @cecheta,

This behavior was a bug that has been fixed. When running a workflow as an agent via .as_agent(), WorkflowOutputEvent events from ctx.yield_output() were not being converted to AgentRunResponseUpdate responses, they were silently ignored.

The fix (now on main and part of 1.0.0b251216) properly converts WorkflowOutputEvent to agent responses in _convert_workflow_event_to_agent_update. With the fix:

workflow_agent = workflow.as_agent()
thread = workflow_agent.get_new_thread()
res = await workflow_agent.run(prompt, thread=thread)

# res.text now contains all yield_output() values
# res.messages contains each output as a separate ChatMessage

For streaming:

async for update in workflow_agent.run_stream(prompt, thread=thread):
    print(update.text)  # Each yield_output() surfaces as an update

Some other notes:

• output_response=True on register_agent() yields the agent's AgentRunResponse as a workflow output
• ctx.yield_output() in custom executors yields arbitrary data as workflow output
• Both now properly surface when using .as_agent()

Please update to the latest version to get this fix.

[cecheta]

Hi @moonbox3 , I've just checked I'm on the latest version, but the behaviour you've described isn't what I'm seeing. Here is the code I'm running:

Code

import asyncio

from agent_framework import (
    AgentExecutorRequest,
    ChatAgent,
    ChatMessage,
    WorkflowBuilder,
    WorkflowContext,
    executor,
)
from agent_framework.azure import AzureOpenAIChatClient
from azure.identity import DefaultAzureCredential


@executor
async def start(
    messages: list[ChatMessage], ctx: WorkflowContext[AgentExecutorRequest, str]
) -> None:
    await ctx.yield_output("Start...")

    await ctx.send_message(
        AgentExecutorRequest(
            messages=messages,
            should_respond=True,
        )
    )


def create_writer_agent() -> ChatAgent:
    """Factory function to create a writer agent."""
    chat_client = AzureOpenAIChatClient(credential=DefaultAzureCredential())
    return chat_client.create_agent(
        instructions=(
            "You are an excellent content writer. You create new content and edit contents based on the feedback."
        ),
        name="writer",
    )


def create_reviewer_agent() -> ChatAgent:
    """Factory function to create a reviewer agent."""
    chat_client = AzureOpenAIChatClient(credential=DefaultAzureCredential())
    return chat_client.create_agent(
        instructions=(
            "You are an excellent content reviewer."
            "Provide actionable feedback to the writer about the provided content."
            "Provide the feedback in the most concise manner possible."
        ),
        name="reviewer",
    )


async def main():
    workflow = (
        WorkflowBuilder()
        .register_executor(lambda: start, "start")
        .register_agent(create_writer_agent, "writer")
        .register_agent(create_reviewer_agent, "reviewer", output_response=True)
        .set_start_executor("start")
        .add_edge("start", "writer")
        .add_edge("writer", "reviewer")
        .build()
    )

    prompt = (
        "Create a slogan for a new electric SUV that is affordable and fun to drive."
    )

    workflow_agent = workflow.as_agent()
    thread = workflow_agent.get_new_thread()
    res = await workflow_agent.run(prompt, thread=thread)

    for msg in res.messages:
        print(f"{msg.role}: {msg.text}")


if __name__ == "__main__":
    asyncio.run(main())

And this is the output:

assistant: Start...
assistant: Clear message and aligns with the vehicle’s qualities. Consider adding a stronger affordability cue to emphasize low cost, e.g., “Adventure Within Reach. Fun, Electrified.” Ensure the slogan is unique and memorable.
assistant: **"Plug Into Adventure. Drive Fun, Drive Smart."**
assistant: Clear message and aligns with the vehicle’s qualities. Consider adding a stronger affordability cue to emphasize low cost, e.g., “Adventure Within Reach. Fun, Electrified.” Ensure the slogan is unique and memorable.

It seems like there are there are two messages for the two workflow outputs, but I also get the agent responses from both agent executors, regardless of whether output_response=True or not.

A similar thing happens when streaming:

    async for update in workflow_agent.run_stream(prompt, thread=thread):
        print(update.text)

Output

Start...

**
"
Electric
 Thr
ills
,
 Everyday
 Bills
."
**



Action
able
 Feedback
:
  

This
 is
 a
 good
 start
!
 The
 slogan
 should
 be
 short
,
 memorable
,
 and
 directly
 convey
 both
 affordability
 and
 fun
.
 Try
 to
 emphasize
 both
 benefits
 in
 a
 catchy
 way
.
 Example
:
  

“
Fun
,
 Fully
 Charg
ed
—
Pric
ed
 for
 Everyone
.”
  

Make
 sure
 the
 slogan
 feels
 energetic
 and
 approachable
 for
 a
 broad
 audience
.


Actionable Feedback:  
This is a good start! The slogan should be short, memorable, and directly convey both affordability and fun. Try to emphasize both benefits in a catchy way. Example:  
“Fun, Fully Charged—Priced for Everyone.”  
Make sure the slogan feels energetic and approachable for a broad audience.

Interestingly, if I use AzureOpenAIResponsesClient instead of AzureOpenAIChatClient, then I get three messages rather than four. The agent responses are now concatenated into a single message, separate from the workflow outputs.

assistant: Start...
assistant: Actionable Feedback:  
Highlight the affordability more clearly, perhaps with a direct mention of value or savings. Also, use a more playful or energetic tone to emphasize the "fun to drive" aspect. Consider something like: "Big Thrills, Small Bills—Drive Fun, Drive Electric!"
assistant: "Electrify Your Drive—Fun. Affordable. Unstoppable."Actionable Feedback:  
Highlight the affordability more clearly, perhaps with a direct mention of value or savings. Also, use a more playful or energetic tone to emphasize the "fun to drive" aspect. Consider something like: "Big Thrills, Small Bills—Drive Fun, Drive Electric!"

[moonbox3]

Thanks for this extra context. The bug I was referring to that was previously fixed is different. I now see what you're saying - will work on fixing this.

[moonbox3]

@cecheta I've converted this to an issue to better track it. Please follow along with #2957.