Python: [Bug]: Copilot Studio Agent and AG-UI Protocol Integration

[yegor-klymenchuk]

Description

When integrating Microsoft Studio Copilot agents with AG UI protocol + CopilotKit, the agent run consistently authenticates successfully and returns HTTP 200, but often produces no assistant message content.

In many runs, the stream contains only:

• RUN_STARTED
• RUN_FINISHED

with no TEXT_MESSAGE_* chunks in between.

Occasionally, the agent does respond with TEXT_MESSAGE_START / CONTENT / END, but this behavior is still not correct and inconsistent.

---

Observed Behavior

Case 1 - No assistant response (most frequent)

The agent run starts and finishes successfully, but no assistant message is emitted:

{"type":"RUN_STARTED", ...}
{"type":"RUN_FINISHED", ...}

• No TEXT_MESSAGE_* events are streamed.

Case 2 - Partial / minimal response (inconsistent)

Sometimes the agent responds with a short assistant message:

{"type":"TEXT_MESSAGE_START", "role":"assistant"}
{"type":"TEXT_MESSAGE_CONTENT", "delta":"Generating plan..."}
{"type":"TEXT_MESSAGE_END"}

---

Expected Behavior

• Every successful agent run (RUN_STARTED -> RUN_FINISHED) should reliably emit assistant content when the agent decides to respond.

---

Question

Is this:

1. Expected behavior of Copilot Studio agents when used via AG UI?
2. A known limitation or configuration issue?
3. Or a bug in the AG UI <-> Copilot Studio integration?

Code Sample

import os

from agent_framework.microsoft import CopilotStudioAgent
from agent_framework_ag_ui import AgentFrameworkAgent
from azure.identity import OnBehalfOfCredential
from microsoft_agents.copilotstudio.client import (
    AgentType,
    ConnectionSettings,
    CopilotClient,
    PowerPlatformCloud,
)


environment_id = os.getenv("COPILOTSTUDIOAGENT__ENVIRONMENTID")
schema_name = os.getenv("COPILOTSTUDIOAGENT__SCHEMANAME")


def create_copilot_studio_agent() -> AgentFrameworkAgent:
    try:
        credential = OnBehalfOfCredential(
            tenant_id=os.getenv("AZURE_TENANT_ID"),
            client_id=os.getenv("AZURE_CLIENT_ID"),
            client_secret=os.getenv("AZURE_CLIENT_SECRET"),
            user_assertion=os.getenv("AZURE_USER_ASSERTION"),
        )

        scope = "https://api.powerplatform.com/.default"
        token_response = credential.get_token(scope)
        token = token_response.token

    except Exception as e:
        raise ValueError(
            f"Failed to authenticate with Azure AD:\n  Error: {str(e)}\n"
        ) from e

    try:
        settings = ConnectionSettings(
            environment_id=environment_id,
            agent_identifier=schema_name,
            cloud=PowerPlatformCloud.PROD,
            copilot_agent_type=AgentType.PUBLISHED,
            custom_power_platform_cloud=None,
        )

        copilot_client = CopilotClient(settings=settings, token=token)
        copilot_agent = CopilotStudioAgent(client=copilot_client)

        return AgentFrameworkAgent(
            agent=copilot_agent,
            name="Microsoft Copilot Studio Agent",
            description="Agent powered by Microsoft Copilot Studio with AG-UI integration.",
            require_confirmation=False,
        )

    except Exception as e:
        raise ValueError(
            f"Failed to create Copilot Studio agent:\n  Error: {str(e)}\n"
        ) from e

Error Messages / Stack Traces

Package Versions

agent-framework-ag-ui: "1.0.0b260107"; agent-framework-azure-ai: "1.0.0b260107"; agent-framework-copilotstudio: "1.0.0b260107"

Python Version

Python 3.12

Additional Context

No response

[moonbox3]

Last I saw, the CopilotStudioAgent had "pseudo-streaming" capabilities? Meaning, we wrapped the normal get response call to return an AsyncGenerator so it seemed like it was streaming. I don't think we'll get reliable usage from a CopilotStudioAgent when running with AG-UI, which expects agents to actually use streaming responses.

[yegor-klymenchuk]

Last I saw, the CopilotStudioAgent had "pseudo-streaming" capabilities? Meaning, we wrapped the normal get response call to return an AsyncGenerator so it seemed like it was streaming. I don't think we'll get reliable usage from a CopilotStudioAgent when running with AG-UI, which expects agents to actually use streaming responses.

Thanks for the clarification - that helps explain the inconsistent behavior.

Given that CopilotStudioAgent currently relies on pseudo-streaming rather than true streaming, I want to better understand what options are realistically available right now.

Questions

1. Supported Usage Path

• Is CopilotStudioAgent officially supported only in non-streaming / request-response scenarios today?
• If so, does that mean AG-UI is fundamentally incompatible with Copilot Studio agents in their current form?

2. Partial Integration Options

• Is there a recommended way to adapt CopilotStudioAgent so that AG-UI receives at least one deterministic assistant message per run?
  • For example, buffering the final response and emitting a single TEXT_MESSAGE block
• Would wrapping the agent to manually emit TEXT_MESSAGE_START / CONTENT / END after RUN_FINISHED be considered a reasonable workaround?

3. Agent-Side Configuration

• Are there Copilot Studio agent settings that affect whether a response is generated at all? e.g conversation mode, response type, publishing configuration, etc

4. Roadmap

• Is true streaming support for Copilot Studio agents planned?
• If so, is there an approximate timeline or a preferred integration pattern we should align with?

---

What I'm Trying to Figure Out

Ultimately, I'm trying to determine whether:

1. We should avoid AG-UI for Copilot Studio agents entirely for now, or
2. There's a supported (or semi-supported) workaround that would allow predictable experience - even if that means single-message responses only

Thanks so much for your help!