Bug: Agent responses lose structured JSON metadata in multi-agent orchestration (MAF 1.x.x)

[giles17]

Bug Report

Summary

Agent responses are sometimes returned in the wrong format. Instead of including both the main text and additional structured data at the end, the system occasionally returns only plain text. This issue appears when handling certain types of interactions and leads to inconsistent output compared to earlier versions, where formatting was always correct and consistent. For details are captured by the Expected Behaviour and Actual Behaviour sections

Steps to Reproduce

Providing exact steps to reproduce this issue is challenging without access to our internal repository and implementation details. The behaviour depends on a specific multi-agent setup, orchestration logic, and prompt configurations that are not publicly available.

Additionally, there have been significant architectural and behavioural changes between the pre-release and the current released version of Microsoft Agent Framework (MAF), which may contribute to the issue.

Expected Behaviour

Observed in:

Package	Version
agent-framework-azure-ai	1.0.0b251204
agent-framework-core	1.0.0b251204

All responses should consistently follow the defined output format, regardless of how the request is handled. Each response must include a plaintext answer followed by the required structured JSON metadata appended at the end.

The use of different agents or interaction types should not affect or alter this formatting.
Correct behaviour:

Actual Behaviour in MAF 1.x.x

When making requests to the agent through the orchestrator, we expect a plaintext response, with JSON metadata appended at the end of the output.

However, when the request is handled by the conversation agent, other sub-agents begin to ignore the output format defined in their prompts. Instead, they start mimicking the first response format, producing only a simple plaintext answer without the expected JSON metadata.

Environment

• OS: Windows 11, WSL with Ubuntu 24.04
• IDE: VS Code 1.116.0

Impact

Medium (workaround exists)

Possible Fix (Optional)

The workaround is to use structured output: https://learn.microsoft.com/en-us/agent-framework/agents/structured-outputs

Reproducibility

Always

[moonbox3]

@giles17 some thoughts around this:

The symptom that the issue is describing is consistent with how our Group Chat and Handoff orchestrations work today. The full conversation history, including the conversation agent's prior plaintext turn, gets broadcast to each sub-agent as input on its next run. LLMs are pretty prone to mimicking format patterns they see in-context even when their system prompt says something different, so a dominant plaintext turn upstream can pull a sub-agent off its "plaintext + JSON suffix" contract. That's prompt-format drift in the model, not the framework dropping or rewriting anything.

The suggested workaround (structured outputs) is honestly the right fix when the format has to be guaranteed. response_format with a Pydantic schema or JSON schema enforces the shape at the model layer and bypasses the mimicry problem entirely. Prompt-only format contracts are a best-effort thing across any agent framework, not something we can enforce from our side.

Before we decide whether there's a real regression here versus working-as-designed, a few things would help:

1. Which orchestration pattern is used? Group Chat, Handoff, Magentic, or something custom on top of workflows?
2. Where is the "plaintext + JSON metadata" contract defined for the sub-agents? System prompt, user prompt, or something else?
3. The issue mentions that the workaround is structured outputs (response_format on the affected sub-agents). Does that fully resolve it?
4. The "earlier versions were always correct" piece is hard to chase without a repro, since model output is non-deterministic and prompt-sensitive. Any chance we can get a minimal repro on a public model with a toy prompt (doesn't have to be anything proprietary)? Even one logged trace showing the orchestrator turn followed by the sub-agent dropping its suffix would be enough to dig in.
5. If you log the messages each sub-agent receives, is the orchestrator's plaintext turn in there? That would confirm the mimicry path vs. something else going on.

Thanks.