Agent response getting truncated in Test panel

(1) Share

Report

Posted on by Nidit Mehta

Copilot Studio: Old UI (Standard harness)

LLM used: Claude Sonnet 4.6

Knowledge: Uploaded file into SharePoint and added SharePoint folder in knowledge of Copilot studio (approx. size of knowledge folder - 1.5 GB, under limited folder and size of individual files)

Agent output format: Agent analyze the user request (behavior driven by agent instructions and grounded response are generated through Copilot Studio's defautl RAG) as per user uploaded file and generates the output in structured tabular format

Agent instructions are under 7000 characters and User prompt under 100 characters

Issue:

- Agent response getting truncated, In the subsequent prompt if I hit "." or "continue" or anything, it provides me the response in the next turn -- That looks like LLM has generated the output in the single turn, but Copilot studio / Test panel isn't displaying the complete output in the single turn

Is there any limitation of Copilot studio or specific to Test Panel in terms of token / context window?

Categories:

General topics

All responses (6)

Answers (0)

Sort by

Sam_Fawzi 1,150 Super User 2026 Season 1 on at

Like
a
(0)

Report
Copy link

Link copied!
Hi @Nidit Mehta

I would question the diagnosis slightly. If the Test panel were hiding text the model had already produced, the fix would be a rendering one. What you describe looks more like an output ceiling on generation, where the model stops mid-answer and then continues from where it left off when prompted, because it can see its own truncated message in the conversation history.

The test that separates the two. Publish the agent and run the identical prompt in a real channel, Teams or the demo website. If it renders completely there, it is a Test panel display issue. If it truncates in the same place, it is generation. Two minutes, and it decides everything that follows.
One more signal worth checking: does it cut off mid-row or mid-cell? A hard stop inside a table row points at a token limit rather than the model choosing to end.

If it is generation, and I suspect it is
Tabular output is expensive. Markdown tables burn a lot of output tokens for the information they carry, all the pipes, separators, and repeated column content. The same data as prose or as JSON costs considerably less. So a structured table is the format most likely to hit a ceiling first.

Things that help, in order of effort:

Reduce columns before reducing rows. Width costs more than you would expect.

Paginate deliberately rather than accidentally. Ask for a fixed number of rows per turn and have the agent state where it stopped. You get the same outcome you are getting now, but predictably and without the user typing a full stop.

Trim your instructions. You are at 7,000 characters against a documented 8,000 character ceiling, and published research shows reasoning quality degrading well before that hard limit. Instructions are not the same budget as output, but tightening them is cheap and worth doing anyway.

One finding from our own work that may be relevant
We hit this on a document extraction pipeline. Reducing input size did not fix it. Reducing the number of fields requested per call did. That is the signature of an output ceiling rather than an input or context problem, and it is worth testing in your case: same prompt, half the columns, see whether it completes.
What ultimately fixed it for us was architectural. We stopped asking the agent to emit all the data and moved the extraction to a deterministic layer, leaving the agent to compose and normalize output that had already been located. If your end goal is structured tabular data rather than conversational answers, having the agent render the table is the expensive way to get there. Returning a compact payload and rendering the table downstream sidesteps the ceiling entirely.

There is no published figure for the output ceiling that I have been able to find, so this is empirical on both sides. Worth posting back what the published channel test shows, since that would be useful to others.
Suggested answer

Sayali-MSFT Microsoft Employee on at

Like
a
(0)

Report
Copy link

Link copied!
Hello Nidit Mehta ,

It looks more like a Copilot Studio response/output limitation rather than the Claude context-window limit.
Copilot Studio doesn’t document a separate fixed token limit specifically for the Test panel, but Microsoft notes that the instructions, retrieved knowledge, conversation context, and generated response must fit within the model’s allowable window. Large outputs can therefore hit token-window or execution limits.

Since“Continue” returns the remaining content, the model is likely able to continue generating, but the full table isn’t being returned in a single turn.

I’d suggest:

Ask the agent to return fewer rows, for example 10–20 at a time.

Test the same query in the published Teams/M365 channel to see if it’s Test-panel-specific.

Save a Test panel conversation snapshot and compare the behavior.

If the response consistently truncates at roughly the same point across channels, raise a Microsoft support case with the snapshot/session details.

Also, the 1.5-GB SharePoint folder itself isn’t passed to the LLM in one prompt. Copilot Studio retrieves relevant content from the configured knowledge source before generating the response.

Reference document:
1. Prompts performance and execution - Microsoft Copilot Studio | Microsoft Learn
2. Quotas and limits - Microsoft Copilot Studio | Microsoft Learn
3. Quotas and limits - Microsoft Copilot Studio | Microsoft Learn
4. Test your agent - Microsoft Copilot Studio | Microsoft Learn
Suggested answer

11manish 4,802 Super User 2026 Season 2 on at

Like
a
(0)

Report
Copy link

Link copied!

The 7,000-character instructions and <100-character prompt are within Microsoft's documented limits. Claude Sonnet 4.6 has a 200K-token context window. The 1.5 GB SharePoint knowledge repository is not itself injected into the model context. Therefore, the most likely issue is the size/complexity of the generated structured response or the interaction between RAG, orchestration, model generation, and response handling—not simply that your agent exceeded the Copilot Studio context window.
Suggested answer

Valantis 7,545 Super User 2026 Season 2 on at

Like
a
(1)

Report
Copy link

Link copied!

Hi @Nidit Mehta,

Checked for a published output length limit for generative answers specifically, found nothing either, confirms what's already been said, there's no documented figure for this.

Two related but different size limits do exist and are documented, worth knowing even if probably not this. Channel data transfer, Direct Line, Facebook, has a hard 262,144 byte request content limit, and conversation state has its own separate size cap that throws QuotaExceeded when hit.https://learn.microsoft.com/en-us/troubleshoot/power-platform/copilot-studio/authoring/error-codes

Both of those come with explicit error messages though, not silent truncation with a working continue, so they're unlikely to be what's happening here. Publishing to a real channel with the same prompt is still the cleanest way to separate Test panel rendering from an actual generation ceiling, since nothing in the docs settles that distinction directly.

Best regards,

Valantis

✅ If this helped solve your issue, please Accept as Solution so others can find it quickly.

❤️ If it didn’t fully solve it but was still useful, please click “Yes” on “Was this reply helpful?” or leave a Like :).

🏷️ For follow-ups  @Valantis.

📝 https://valantisond365.com/

💼 LinkedIn

▶️ YouTube
Suggested answer

Mohsin Ali 1,075 on at

Like
a
(0)

Report
Copy link

Link copied!

Hello @Nidit Mehta - In the standard harness, Copilot Studio can make multiple model calls within a conversational turn, and Microsoft mentions that the total token usage for a turn can exceed the context window of an individual model call. You may also want to check the documentation below for more details.

Since typing "continue" returns the remaining output, it looks more like the response is being split or truncated at the Copilot Studio harness or presentation layer, rather than the model failing to generate the complete answer.

I would suggest testing the same prompt with a smaller output, for example by limiting the table to 20 or 50 rows. Also, compare the behavior in the Test pane with a published channel. If smaller responses work consistently, that would point more toward an output-size or harness limitation.

Also, your ~7,000-character agent instructions are within Microsoft's documented 8,000-character limit.

I’m also adding the Microsoft documentation references below. They are worth reviewing for a better understanding of model settings, context windows, and Copilot Studio limits:

https://learn.microsoft.com/en-us/microsoft-copilot-studio/prompt-model-settings

https://learn.microsoft.com/en-us/microsoft-copilot-studio/requirements-quotas
Suggested answer

M Bilal Khan 376 on at

Like
a
(0)

Report
Copy link

Link copied!

I’ve seen similar behavior in Copilot Studio, especially when the response is large and being generated in a structured/table format.

The fact that sending “continue” immediately produces the remaining content is a good indication that the model may have generated more content, but the response is being cut off at the Copilot Studio/test harness side rather than the knowledge retrieval itself.

I’d suggest testing the same prompt with a much smaller output first (for example, limit the number of rows/sections returned) and also testing it outside the Test panel if possible. This can help determine whether the truncation is related to the test harness or to the agent/runtime response limits.

Also worth checking the generated response in the Copilot Studio conversation/transcript and monitoring the token usage/response details if available. If the issue consistently occurs only with large tabular responses, I would consider breaking the output into smaller sections and explicitly instructing the agent to return a maximum number of rows/items per turn.

So I wouldn’t assume the 1.5 GB knowledge size is the direct cause here. The more relevant area to investigate would be the maximum response/output size and how the Test panel handles long generated responses.

If anyone has reproduced this specifically with Claude Sonnet 4.6 in the current Copilot Studio experience, it would be useful to compare the same prompt in the published agent versus the Test panel.