Building a semi-development agent

1. Goal of this chapter

This chapter walks through the process of creating an actual agent with Copilot Studio. From conversation design → variable management → action connections → Generative Orchestration, it shares a practical methodology validated in field workshops.

2. Choosing a conversation design approach

There are two main ways to design an agent’s conversation in Copilot Studio. They are not mutually exclusive; you can freely combine them within a single agent.

2.1 Instructions-based design

When you write instructions in natural language on the agent’s Overview tab, the LLM autonomously leads the conversation based on them.

Suitable when:

• Questions are open-ended and varied
• The flow should follow “knowledge → judgment → action”
• B2E (internal employee-facing) work support, or mixed knowledge + automation scenarios

Recommended instruction structure (practical template):

## 3. Role
You are an expert [business domain] agent for [organization name].

## 4. Order of behavior
1. First classify the intent of the question
2. If it is unclear, ask follow-up questions
3. If you can answer, answer based on Knowledge
4. If work is needed, call the appropriate Action/Flow
5. Summarize the result and explain it to the user

## 5. Constraints
- Provide only answers grounded in internal policy documents
- If there is no basis, say "confirmation is needed"
- Before sending an email/message, always ask the user to confirm the content

## 6. Available tools
- [HR policy search] — SharePoint Knowledge
- [Send email] — Outlook connector
- [Create IT ticket] — ServiceNow Flow

📖 Reference: Generative orchestration guidance

6.1 Topic-based design

When structured business processing is needed, design the conversation in the Topics tab as a trigger → node flow.

Suitable when:

• The work procedure is fixed (help desk, approval process)
• Branches/conditions are clear
• It is an “area that must not be wrong”

Topic design flow:

Define trigger phrases (or automatically match by topic name/description in Generative mode)
  → Question node (collect required information → store in variables)
    → Condition node (branch based on variable values)
      → Action node (call external system)
        → Return result through output variables (message node ❌)

📖 Reference: Create and edit topics

6.2 Which approach should you choose? — decision criteria

7. How to use variables

Use the Variable system to store and pass data during conversations.

7.1 Variable types

📖 Reference: Work with variables · Variables overview

7.2 Common variable patterns

7.3 Key system variables

📖 Reference: System variables list

7.4 Variable mapping order (practical must-follow)

When using variables, you must follow this order. If you ignore the order, values can be empty at runtime.

1. Create the variable first (define name and type)
2. System / Topic / Prompt output → store in variable
3. Variable → map to Flow input
4. Flow output → receive back into variable
5. Use variable values in messages/conditions

7.5 Passing variables between topics

When passing data between topics, the Redirect + return variable pattern is recommended instead of global variables.

1. Sending topic: pass values from the Redirect node
2. Receiving topic: check “Receive values from other topics” in variable properties
3. Return: check “Return values to original topics” in variable properties

This lets you create clean data flows without overusing global variables.

📖 Reference: Pass variables between topics

7.6 Initializing variables from external sources

When embedding an agent in a website, you can pass user information in advance.

1. Create a global variable → check “External sources can set values”
2. Pass via URL query parameter: botURL?UserName=AlexKim

📖 Reference: Set global variables from external sources

7.7 Using Power Fx formulas

Copilot Studio supports the Power Fx formula language to handle complex operations and data transformations.

// Concatenate strings
Concatenate(VarFirstName, " ", VarLastName)

// Conditional message
If(VarRequestType = "Urgent", "Forwarding to the urgent handling team.", "Received as a standard request.")

// Date calculation
DateAdd(Now(), 3, TimeUnit.Days)

📖 Reference: Create expressions using Power Fx

8. Action-based processing flow

An Action is the core capability for an agent to call external systems during a conversation and perform real work. It is the point where an “answering Copilot” turns into a “working agent.”

8.1 Comparing action connection methods

8.2 Practical tips for designing Power Automate flows

Recommended architecture:

Copilot Studio Agent
  └─ Action call (Agent handles judgment/branching)
       └─ Power Automate Cloud Flow (executes only structured work)
            └─ (if needed) Desktop Flow / external API / MCP

Core design principles:

8.3 Connector-based action setup (step by step)

1. Click the Actions tab → Add an action
2. Search for the connector to use (for example, “Office 365 Outlook”)
3. Select an operation (for example, “Send an email (V2)”)
4. Set input parameters to variables or fixed values
5. Select an authentication method

📖 Reference: Use actions with custom agents

9. Practical guide to Generative Orchestration

9.1 How to enable it

1. Select the Settings button → Generative AI
2. Select the Generative radio button (default: Classic)
3. Set the content moderation level
4. Click Save

📖 Reference: Orchestrate agent behavior with generative AI

9.2 Core design points in Generative mode

In Generative mode, topic/action names and descriptions are more important than trigger phrases. The LLM uses them to decide which tool to use.

Example of a good topic name/description:

Name: IT equipment request
Description: Used when employees request IT equipment such as laptops, monitors, and keyboards.
             Collects the required equipment type and purpose of use, then registers them in SharePoint.

Example of a good action description:

Name: Send email (Outlook)
Description: Sends an email to the specified recipient.
             Calls this only when requested by the user, and always confirms the content before sending.

9.3 Testing and debugging

• Activity Map: When enabled in the test panel, you can visually trace which topic the agent selected and which action it called.
• Variable watch: In the Test tab on the side of the test panel, you can check current variable values in real time.

10. Top 7 mistakes from workshops

These are mistake patterns repeatedly observed in field workshops. Avoiding just these significantly improves agent quality.

11. Summary of agent design best practices