M365 Copilot Agent does not update correctly

Linke, Jonas (I/VH-45) 0 Reputation points
2026-05-19T11:23:15.9433333+00:00

Hi all,

Ive built many M365 Copilot Premium Agents and am constantly updating them to finetune and update.

Now the updated prompt is saved BUT the agent does use the "old prompt" version. Also the time stamp "last updated" remains unchanged (e.g. last updated April 16th --> but should be last updated May 19th).

How can I troubleshoot and correct this behavior? Ive deleted my chache, site and cookies... but without change.

Thanks for support!

Microsoft Copilot | Microsoft 365 Copilot | Development
0 comments No comments

2 answers

Sort by: Most helpful
  1. Sayali-MSFT 6,396 Reputation points Microsoft External Staff Moderator
    2026-05-20T10:48:22.99+00:00

    Hello Linke, Jonas (I/VH-45),
    The issue you’re facing is typically caused by Copilot Studio using a cached or previously published version of your agent rather than the latest saved draft. Even though your prompt changes are saved, they will not take effect until the agent is properly published and reloaded across all client surfaces like Teams or Microsoft 365 Copilot. Additionally, existing chat sessions often continue using the older configuration, which is why your agent behaves as if the old prompt is still active.

    The unchanged “last updated” timestamp further indicates a publish or propagation delay issue, where updates have not fully synced across the Copilot service and connected channels. This is common due to backend caching, session persistence, and delayed distribution across regions. To resolve it, you typically need to publish again, start a completely new chat session, and in some cases remove and re-add the channel or reinstall the Teams app to force a refresh of the agent configuration.

    Was this answer helpful?

    0 comments No comments

  2. AI answer

    2026-05-19T11:23:33.2933333+00:00

    The behavior matches known caching and propagation delays for Microsoft 365 Copilot agents and their suggested prompts, not a failure to save.

    To troubleshoot and correct:

    1. Confirm the agent is actually saved and published
      • In Copilot Studio, after editing the agent (system message, instructions, knowledge, suggested prompts, etc.), select Save and then Publish.
      • If the agent isn’t published, Copilot will continue to use the previously published configuration and the “last updated” timestamp in the Copilot surfaces will not change.
    2. Start a completely new session
      • Existing chats can keep using the old configuration.
      • Close all existing chats with the agent and start a brand‑new chat session with it (for example, in Microsoft 365 Copilot web or Teams) so the new configuration is loaded.
    3. Clear Teams/M365 Copilot caches and reattach the agent If using the agent in Teams or Microsoft 365 Copilot:
      • Clear the Teams cache (if applicable).
      • Remove the agent’s channel/app from Teams, then add it back.
      • Publish the agent again and allow time for global synchronization. These steps help when the client or service is still serving cached metadata (including prompts and timestamps).
    4. Allow for propagation time
      • Updates to suggested prompts and agent configuration might not appear immediately after publishing because of caching in the browser, Teams/Microsoft 365 Copilot services, and the content delivery network.
      • Publishing usually succeeds, but it can take time for changes (including “last updated”) to propagate across all regions.
    5. Use versioning and validate in M365 Copilot web
      • Use versioning for prompt/agent updates so it’s clear which version is live.
      • Validate changes first in the Microsoft 365 Copilot web experience, which often reflects updates sooner and helps distinguish a client‑side cache issue from a publishing issue.

    If, after confirming save + publish, starting a new session, clearing caches, and waiting for propagation, the agent still uses the old prompt and the timestamp never advances, capture the agent ID, the expected vs. actual behavior, and open a support ticket so the backend configuration and propagation pipeline can be checked.


    References:

    AI-generated content may be incorrect. Read our transparency notes for more information.

    Was this answer helpful?

Your answer

Answers can be marked as 'Accepted' by the question author and 'Recommended' by moderators, which helps users know the answer solved the author's problem.