AI Labs GitHub
Table of contents

Lab 03 - Capabilities, skills and the jump to Copilot Studio

Lab10 min readLab time: 40 min

Lab introduction

So far our agent can talk and read. Now we let it work.

Agent Builder gives you three ways to add muscle without writing code:

  • Code interpreter ("Create documents, charts, and code") - the agent writes and runs Python in a sandbox to analyze data, build charts and create Word, Excel and PowerPoint files.
  • Image generator ("Create images") - the agent creates images from a prompt.
  • Skills - a packaged, repeatable procedure (a SKILL.md file plus optional scripts and resources) that the agent loads every time a matching task comes up.

Then we hit the ceiling. Agent Builder can't call external systems - it can't create a SharePoint item, send an email or start an approval. For that we need Copilot Studio. So at the end of this lab we carry the agent across.

Lab scenario

The Service Desk manager sends a CSV export of Q3 tickets every quarter and asks the same questions: which category is growing, where do we miss the SLA, which site is unhappy? She wants the IT Helper to answer these in seconds, the same way every time. She also wants every incoming request triaged with the same rules - category, priority, team.

Job skills

  • Task 1: Analyze data with code interpreter
  • Task 2: Create a skill from a description
  • Task 3: Upload a skill package
  • Task 4: Hit the ceiling
  • Task 5: Copy the agent to Copilot Studio

Task 1: Analyze data with code interpreter

  1. Edit Contoso IT Helper (Agents > ... > Edit). In the Knowledge section, select the ⚙️ gear. In Knowledge settings, check that Create documents, charts, and code is on. Close the dialog.

Knowledge settings with Create documents, charts, and code on

  1. Download Allfiles/Labs/03/Contoso-IT-Tickets-Q3-2026.csv from this repository. It has 240 tickets with these columns: TicketID, OpenedDate, Site, Category, Priority, ResolutionHours, WithinSLA, SatisfactionScore.

The ticket export in the GitHub repository

  1. Open the Configure view menu at the top of Agent Builder. Notice the third option, Monitor, next to Configure and Preview - it shows how your agent is used after you share it. For this task, close Agent Builder and open the agent's own chat (Agents > Contoso IT Helper), so the generated files open full size.

View menu with Configure, Preview and Monitor

  1. Select + in the message box and attach the CSV.

CSV attached in the agent chat

  1. Type this prompt and send it:
Text
Using this ticket export, give me:
1. the number of tickets per category, as a bar chart,
2. the SLA compliance rate (WithinSLA = Yes) per priority,
3. the site with the lowest average SatisfactionScore.
Keep the commentary to five bullet points.

Analysis prompt with the CSV attached

  1. Watch the status: Contoso IT Helper: Coding and executing. Select the arrow to see the Python code it writes and runs.

Coding and executing status

  1. Review the result: - a bar chart of tickets per category (Software 73, Hardware 61, Network 41, Account & Access 40, Security 25),

    Tickets per category chart

  • an SLA compliance table per priority, and five bullets citing the CSV as the data source.

    SLA compliance by priority

  • the site with the lowest satisfaction: Madrid, 3.38 out of 5.

  1. Now ask for a file:
Text
Create an Excel file with one sheet per category and a summary sheet

The workbook opens next to the chat: a Summary sheet with the three tables and a chart, plus one sheet per category.

Excel workbook with a Summary sheet and one sheet per category

Check yourself

Question: Why use code interpreter instead of just asking the model to "count"? Answer: Language models are bad at arithmetic over 240 rows. Code is exact. When numbers matter, make the agent compute, not guess.


Task 2: Create a skill from a description

Every quarter the manager asks for the same report. Typing the same five-line prompt each time is the definition of a skill.

  1. Edit the agent again. In the Agent Builder chat (left pane), paste:
Text
Add a reusable skill called "quarterly-sla-report".
When the user uploads a Contoso ticket export (CSV with TicketID, OpenedDate, Site,
Category, Priority, ResolutionHours, WithinSLA, SatisfactionScore), always:
1. compute tickets per category and per site,
2. compute SLA compliance per priority and flag any priority below 85%,
3. find the 3 slowest categories by median ResolutionHours,
4. produce a one-page Word report with a title, a 5-bullet executive summary,
   the tables and one bar chart.
Package it with clear instructions and the analysis script so the result looks the same every time.
  1. Scroll to Skills. You should see quarterly-sla-report. In our run, Agent Builder also created a second skill, ticket-export-analysis, for the ad-hoc analysis you did in Task 1 - and added lines to the instructions saying when to invoke each one.

Skills list with it-incident-intake, quarterly-sla-report and ticket-export-analysis

Important

Agent Builder may create extra skills on its own. That's fine - but read each one (select the > arrow) and delete what you don't want. You own the agent, not the generator. A skill limit applies: eight skills per agent.

  1. Select Update.

  2. Open the agent chat, attach the CSV again and type only:

Text
Quarterly report please

Quarterly report please with the CSV attached

  1. The agent builds Contoso-Q3-2026-Quarterly-SLA-Report.docx: title, 5-bullet executive summary, tickets by category and by site, SLA compliance by priority with a Flag column.

The generated Word report

  1. Run it a second time in a new chat. Same structure? That's the point of a skill.

Note

Look at the Flag column. With our sample data, every priority is below the 85% target, so every row says "Below 85%". Correct - but is that useful? Discuss: should the skill also show how far below target each priority is? Edit the skill (Agent Builder chat: "In quarterly-sla-report, add the gap to 85% for each priority") and run it again.


Task 3: Upload a skill package

Skills are portable. The same package works in Agent Builder, in Copilot Studio (Lab 07), and in other tools that support the open Agent Skills format. Let's prove it.

  1. Open Allfiles/Labs/03/it-ticket-triage/SKILL.md in any text editor and read it. Notice the structure: - a YAML front matter with name and description (the description tells the agent when to use the skill), - the rules (category signals, priority rules, team mapping), - a strict output format (Category / Priority / AssignedTeam / Reason).

  2. In Agent Builder, scroll to Skills and select + Add. The Upload skill dialog accepts a SKILL.md file or a .zip package that includes a SKILL.md. Upload it-ticket-triage/SKILL.md (or it-ticket-triage.zip) and select Upload.

Upload skill dialog with SKILL.md

Note

Use the .zip when the skill has extra files (scripts, templates). A lone SKILL.md is enough for an instructions-only skill like this one.

  1. Keep the window open while the skill uploads.

Uploading skill

  1. Switch to Preview, select New chat, and test:
Message Expected
Triage this: "I clicked a link in an email and now my browser opens strange pages" Security issue - the urgent instruction (call ext. 4357), no ticket
Triage this: "Need Power BI Pro for the sales dashboard, no rush" Software, Low, Apps
Triage this: "Left my laptop in a taxi in Madrid 20 minutes ago" Urgent - report within 4 hours, Intune wipe
  1. Test 1 - notice that the Boundaries you wrote in Lab 01 win over the skill: for a security problem the agent gives the urgent instruction and stops, instead of producing a triage block.

Phishing: security instruction, no ticket

  1. Test 2 - the exact output format from the skill.

Power BI triaged as Software, Low, Apps

  1. Test 3 - the agent treats the lost laptop as urgent and cites the Laptop and Device Policy.

Lost laptop in a taxi handled as urgent

  1. Select Update to save the agent.

Task 4: Hit the ceiling

  1. In the same Preview chat, type:
Text
Please create the ticket in the IT Requests list for me
  1. Read the answer: "I couldn't create the item because no tool is available to write to the IT Requests list." The agent writes the ticket content for you to copy - and still tells you to call the Service Desk now.

The agent can't write to the IT Requests list

That's the ceiling. Agent Builder doesn't author actions that call external services.

Spot it

Spot the mistake. Look at the ticket in the screenshot: Category: Security. Our triage skill says a lost or stolen device is Hardware, High - Security is for phishing, suspicious sign-ins and data leaks. Two instructions pulled in different directions (the lost-device urgency vs. the triage rules). Fix it: in the Agent Builder chat, type "A lost or stolen device is always Category Hardware, Priority High. Security is only for phishing, suspicious sign-ins, malware or data leaks." Select Update, then re-run test 3 and this test. This is exactly why we test: two good rules can still conflict.


Task 5: Copy the agent to Copilot Studio

Time to move up.

  1. In Agent Builder, open ... (top-right, next to Update). Look at the options:
Option What it does
Share Share with colleagues (Lab 02)
Agent details Name, icon, description
Submit to org catalog Ask the admin to publish the agent for the whole organization
Download .zip file Export the agent package
Copy to Copilot Studio Our next step

More options menu with Copy to Copilot Studio

  1. Select Copy to Copilot Studio. Read the dialog Copy to full version of Copilot Studio?: - Comes over: name, icon, description, instructions, suggested prompts, enterprise knowledge, SharePoint and website sources. - Not included: uploaded files, Teams and Outlook knowledge sources, sharing settings, image generation, and actions.

Copy to full version of Copilot Studio dialog

Note

Notice what's not listed: skills. Plan to re-add your skills in Copilot Studio (we do it in Lab 07).

  1. Select Get started. In Choose an environment, pick the environment you wrote down in Lab 00 (here MTS Team (default)) and select Continue.

Choose an environment

Note

The list shows every environment you can build in - your personal developer environment may appear too. Pick the one with Dataverse that your trainer chose for the course.

  1. Copilot Studio opens in a new tab on Start building your agent, with Contoso IT Helper (copy): details, description, instructions and suggested prompts are already filled in. Notice the message Knowledge is being prepared. Select Create.

Start building your agent with the copied configuration

  1. The agent opens on its Overview page with status Setting up.... Wait until setup finishes before testing.

Copied agent overview - setting up

  1. Look at the tabs across the top: Overview, Knowledge, Tools, Agents, Topics, Activity, Evaluation... This is the classic Copilot Studio designer - the standard harness, with topics. In Lab 04 you create a new agent from the Agent tile on the Home page, which runs on the Microsoft Copilot harness with a single Build tab. Compare the two experiences.

Note

Important: a copy is a snapshot, not a sync. Changes you make to the Agent Builder version later won't reach the Copilot Studio version, and the other way round. Your original agent keeps working in Microsoft 365 Copilot.

We keep this copy as a reference. In Lab 04 we build the "real" Service Desk agent in Copilot Studio from scratch, so you understand every piece.


Check your work

  • The agent produced a bar chart and an Excel workbook from the CSV.
  • quarterly-sla-report produced a Word report, and it-ticket-triage returns Category / Priority / AssignedTeam / Reason.
  • A lost device is triaged as Hardware, not Security.
  • Contoso IT Helper (copy) exists in Copilot Studio.

Key takeaways

  • Code interpreter makes the agent compute instead of guess - and create real Excel and Word files.
  • A skill = a repeatable procedure the agent loads only when the task matches its description. Agent Builder may generate skills for you: review them.
  • Skills upload as a SKILL.md or a .zip - and travel between tools.
  • Instructions and skills can conflict. Test, spot it, fix it.
  • Agent Builder stops at actions. Copy to Copilot Studio brings the agent to the classic (standard harness) designer; uploaded files, Outlook/Teams knowledge, images and skills need to be re-done.

Next: Lab 04 - Your first Copilot Studio agent

View the source on GitHub