How can I make my Copilot Studio Agent give consistent response?

(1) Share

Report

Posted on by CU12031506-1

I’m building a Copilot Studio agent that fetches meetings from the user’s Outlook calendar and then calculates:

1. total meetings for today
2. conflicts (overlapping meetings)
3. back‑to‑back meeting blocks
4. focus time inside office hours (9 AM–6 PM)

Even though I’ve given detailed instructions in the system prompt, the agent sometimes gives different answers for the same email calander, especially when calculating overlapping meetings or partial overlaps inside office hours.

My questions are:

1. How can I make the agent produce consistent, deterministic results every time?
2. Is there a recommended way to structure prompts so the agent always counts meetings, conflicts, and overlaps correctly?

Categories:

Calling actions from Copilot Studio

General topics

I have the same question (0)

All responses (3)

Answers (0)

Sort by

lbendlin 8,716 Super User 2026 Season 1 on at

Like
a
(1)

Report
Copy link

Link copied!

Copilot (and any other LLM based tool) is probabilistic by design. There is absolutely no guarantee that you will get consistent responses. If you need consistency, use a programming language.

Was this reply helpful? Yes No
Suggested answer

DP_Prabh 381 on at

Like
a
(1)

Report
Copy link

Link copied!
Hi @CU12031506-1,

In Microsoft Copilot Studio, inconsistent results happen because the LLM performs probabilistic reasoning, so calculations like overlaps can vary. How to get deterministic results

Move the calculation logic to a tool/action (Power Automate flow or API) instead of doing it in the prompt.
Let the tool compute:

total meetings

conflicts

back-to-back blocks

focus time

The agent should only display the results, not calculate them. How to improve prompt consistency (if logic stays in the agent)

Pass meetings as structured data (JSON).

Define exact rules for overlaps (e.g., startB < endA AND endB > startA).

Provide step-by-step instructions (sort meetings → check overlaps → compute time in minutes).

Require a fixed output format.

Best practice: Use a Power Automate / tool for calculations and keep the agent responsible only for summarizing results.

I hope this works for you!

Was this reply helpful? Yes No
Assisted by AI

Prasad-MSFT Microsoft Employee on at

Like
a
(1)

Report
Copy link

Link copied!

1. How to Make the Agent Produce Consistent, Deterministic Results
Preprocess Data:
Instead of letting the agent interpret raw calendar data, preprocess the Outlook events in your workflow (Power Automate, code, or a pre-step) to:

Sort meetings by start time.
Normalize time zones.
Filter only today’s meetings before passing to the agent.
Use Structured Input:
Pass meetings as a structured list (e.g., JSON array with start/end times) rather than free text. This reduces ambiguity.

Explicit Instructions:
In your system prompt, specify:

“Do not guess. Only use the provided meeting data.”
“Always use 24-hour time format.”
“If two meetings overlap by even one minute, count as a conflict.”
Set Temperature to 0:
If you have control over model parameters, set temperature to 0 for maximum determinism.

2. Recommended Prompt Structure for Accurate Counting
Prompt Example:

You are a scheduling assistant. Given a list of meetings for today, perform the following:
1. Count the total number of meetings.
2. Identify all pairs of meetings that overlap (even partially).
3. Identify blocks of back-to-back meetings (where one ends and the next starts within 5 minutes).
4. Calculate total focus time between 09:00 and 18:00 that is not occupied by meetings.

Rules:
- Meetings are provided as a JSON array with start and end times in ISO 8601 format.
- Only use the provided data. Do not infer or guess.
- For overlaps, if any part of two meetings overlap, count as a conflict.
- For focus time, subtract all meeting times from the 09:00–18:00 window.

Meetings:
{{meetings_json}}

Return your answer as a JSON object with keys: total_meetings, conflicts, back_to_back_blocks, focus_time_minutes.

Tips:

Use clear, step-by-step instructions.
Define what counts as “overlap” and “back-to-back.”
Ask for output in a structured format (JSON) to avoid ambiguity.

3. Additional Tips
Test with Edge Cases:
Include test data with overlapping, adjacent, and partial meetings to validate the agent’s logic.

Was this reply helpful? Yes No