Forum Discussion
Fabric Data Agent published to M365 has lost conversational context since yesterday
Hello,
I’m experiencing a strange issue with a Fabric Data Agent that has been published to Microsoft 365.
The agent continues to work correctly when accessed directly from Fabric. However, when using the same agent through Microsoft 365, it appears to have lost its ability to maintain conversational context and follow-up questions.
Up until Friday, everything was working as expected, and I have not made any changes to the agent configuration, semantic model, capacity, permissions, or data sources.
Example:
User: Give me the payments for July.
Agent: (Returns the correct results.)
User: Give me the payments for June as well.
Agent: "What would you like me to provide?"
Previously, the agent correctly understood that the follow-up question referred to payments and would return the requested information. Now it seems unable to retain context between messages.
The issue only occurs in the Microsoft 365 published version of the agent. The same agent continues to behave correctly inside Fabric.
Has anyone else noticed a similar degradation in conversational context retention or follow-up question handling during the last couple of days?
Any feedback would be appreciated.
Thank you.
Thanks for sharing the detailed explanation. This sounds like an issue with the Microsoft 365 published agent rather than the Fabric Data Agent itself, especially since the same agent continues to maintain context correctly within Fabric.
A few things you could try:
- Republish the agent to Microsoft 365 to ensure the latest configuration is synced.
- Verify that conversational memory or session context is still enabled after publishing.
- Clear any cached sessions in Microsoft 365 and start a new conversation to rule out stale session data.
- Compare the agent version deployed in Microsoft 365 with the one currently running in Fabric
- Review the Microsoft 365 and Fabric service health dashboards to check for any recent platform issues or regressions.
If none of these resolve the problem, it may be a recent service-side regression in the Microsoft 365 integration. I would recommend opening a Microsoft support ticket and sharing the conversation IDs, tenant ID, timestamps, and a reproducible example, as that will help the product team investigate more quickly.
Please let us know if you find the root cause. It would be helpful for others who encounter the same behavior.
5 Replies
- v-csrikanth
Community Support
Hi John_DataYoda
As suggested by imjrj please raise the support ticket so that our internal team will work on to resolve your issue.
Thank you. - v-csrikanth
Thanks for the detailed description.
Since the same Fabric Data Agent continues to maintain conversational context correctly when accessed directly from Fabric, the agent configuration and semantic model appear to be working as expected. The behavior looks isolated to the published Microsoft 365 Copilot experience.
Recommended workaround next steps:- Republish the Data Agent to Microsoft 365 from the Fabric agent page.
- Start a new Microsoft 365 Copilot conversation and retest.
- Compare two follow-up styles to narrow down the trigger:
Explicit follow-up - Give me the payments for June.
Contextual follow-up - Give me the payments for June as well. -
If the issue persists only in Microsoft 365, capture a sample conversation with timestamps (and the working equivalent from Fabric) and open a Microsoft Support ticket from the Fabric admin portal this may need product-team investigation.
If other community members are noticing the same behavior around the same timeframe, please share here that will help determine whether it's a broader service-side issue.
References links:Consume Fabric Data Agents in Microsoft 365 Copilot
Thanks,
Srikanth Cheri
Community Support Team
- v-csrikanth
Hi John_DataYoda
We would like to inquire whether have you got the chance to check the solutions provided above in community to resolve the issue. We hope the information provided helps to clear the query. Should you have any further queries, kindly feel free to contact the Microsoft Fabric community.Hi v-csrikanth ,
Thank you for following up.
Unfortunately, the issue still remains and has been occurring throughout the entire week.
I followed all the recommended steps, including republishing the Data Agent to Microsoft 365 and testing in new conversations, but the behavior did not improve.
As shown in the attached screenshot, the agent published to Microsoft 365 is unable to maintain conversational context. In this example, after asking about payments in Cairo, I simply reply with "2025". Instead of understanding that the year refers to the previous question, the agent treats it as a completely new request and asks what information I want for 2025.
The same Data Agent continues to work correctly inside Fabric and maintains context as expected.
What makes me believe this is not a configuration issue is that I have used this exact conversational pattern hundreds of times in the past, and until recently it worked perfectly in Microsoft 365 as well. No changes have been made to the agent configuration, semantic model, permissions, or data sources.
I have also noticed inconsistent behavior during the day. Occasionally, the agent will correctly maintain context for a short period of time, but shortly afterward it reverts to losing context and becomes unable to remember even the immediately previous message.
Based on my testing, the issue appears to be specific to the Microsoft 365 published experience rather than the Fabric Data Agent itself.
Has there been any recent change in the Microsoft 365 Copilot orchestration or context-handling logic that could explain this behavior?
Thank you.
John
Thanks for sharing the detailed explanation. This sounds like an issue with the Microsoft 365 published agent rather than the Fabric Data Agent itself, especially since the same agent continues to maintain context correctly within Fabric.
A few things you could try:
- Republish the agent to Microsoft 365 to ensure the latest configuration is synced.
- Verify that conversational memory or session context is still enabled after publishing.
- Clear any cached sessions in Microsoft 365 and start a new conversation to rule out stale session data.
- Compare the agent version deployed in Microsoft 365 with the one currently running in Fabric
- Review the Microsoft 365 and Fabric service health dashboards to check for any recent platform issues or regressions.
If none of these resolve the problem, it may be a recent service-side regression in the Microsoft 365 integration. I would recommend opening a Microsoft support ticket and sharing the conversation IDs, tenant ID, timestamps, and a reproducible example, as that will help the product team investigate more quickly.
Please let us know if you find the root cause. It would be helpful for others who encounter the same behavior.
