Answered

Claude Model Outputs tags to user in M365 Chat on Deeper Thinking questions

(0) Share

Report

Posted on by JY-30040148-0

Wondering if anyone has also experienced formatting issues for published agents which are running on Claude models (using Claude Opus 4.5 specifically) in Copilot Studio Full when talking to the agent in M365 Copilot Chat.

I have tried suppressing / formatting the response via the Agent Instructions to use Markdown and to not print tags however the Claude output still responds with formatting artefacts especially for longer detailed responses. This includes full <thinking> blocks and sometimes only <thinking? blocks.

Seems like this is only an issue for Claude Models as I never had an issue with GPT-5. This also does not occur in the published Teams Chat version.

If anyone has insight on how they overcame this it would be greatly appreciated!

Screenshot 2026-0...

Your file is currently under scan for potential threats. Please wait while we review it for any viruses or malicious content.

Categories:

Publish & channel management

I have the same question (0)

All responses (2)

Answers (1)

Sort by

Suggested answer

Sunil Kumar Pashikanti 2,318 Moderator on at

Like
a
(0)

Report
Copy link

Link copied!
Hi @JY-30040148-0,

If you are seeing <thinking> tags in Copilot Studio agents using Claude in M365 Copilot Chat, you are not doing anything wrong. This is a known issue.

What is happening

Claude models sometimes include internal reasoning wrapped in <thinking> tags. Most channels like Teams filter this out, but M365 Copilot Chat does not always clean it properly, especially for longer answers.

Why instructions don’t help

Even if you tell the agent not to use tags, this does not always work. The tags are part of the model’s internal process, and the issue is that the chat UI does not fully remove them after generation.

What you can do

Use GPT models if you need clean output

Keep responses shorter to reduce the issue

Use Teams channel where filtering works better

Takeaway

This is a formatting gap between Claude and M365 Copilot Chat, not a problem with your prompts.

✅ If this answer helped resolve your issue, please mark it as Accepted so it can help others with the same problem.
👍 Feel free to Like the post if you found it useful.

Sunil Kumar Pashikanti, Moderator
Blog: https://sunilpashikanti.com/posts/

Was this reply helpful? Yes No
Verified answer

JY-30040148-0 9 on at

Like
a
(0)

Report
Copy link

Link copied!

Through testing, I have identified that the root cause appears to be the use of sub-agents within a main agent. When sub-agents are invoked, they inherit the Claude model, and their outputs are then passed back to the main agent with additional internal instruction or reasoning tags (for example, <thinking> tags).

This behaviour results in these tags being surfaced in Copilot Studio outputs, which is not expected and causes downstream issues. Based on this behaviour, it appears to be an unintended side effect or bug, rather than designed functionality. Additional guardrails or output sanitisation may be required to ensure smooth and consistent operation when Claude models are used in conjunction with sub-agents.

Was this reply helpful? Yes No