Models, instructions and prompts for agents

Behaviour of a Copilot Studio agent comes mostly from the model and the text you give it. The exam tests where you set each and how you keep answers controlled.

Agent instructions

Instructions are the system-level text for the whole agent. Good instructions are short and specific:

• Role and audience (“You help HR staff with leave policy”).
• Scope and refusals (“Do not answer legal questions”).
• Format and tone (“Answer in three bullets or fewer”).
• Tool and knowledge rules (“Use the Leave Balance tool before answering about days left”).

Refer to tools and knowledge by name so orchestration can connect them. Instructions do not replace descriptions on each tool.

Prompts as tools

A prompt (built in the prompt builder, part of AI Builder) is a reusable instruction with input variables. You can test it, pick its model, and call it from an agent, topic or flow. Typical uses: summarise text, classify a message, extract fields into structured output.

You can add knowledge to a prompt, for example a Dataverse table, so it uses your data.

Models

Option	Use
Default Copilot Studio model	General agents; Microsoft manages it
Other available models	Test for quality, speed or cost
Microsoft Foundry models (formerly Azure AI Foundry)	Bring your own or an approved catalog model; governed use

Model choice affects quality, latency and cost, and some regions or tenants restrict which models are allowed. When a requirement says responses must use an approved model, select that model explicitly in the prompt or agent settings, and publish the agent so the change takes effect in the channel.

Grounding and safety

• Turn on citations so answers can be traced.
• Control whether the agent may use general knowledge or only your sources. Disabling it keeps answers grounded.
• Content moderation settings control how strict the model is with harmful content.
• Do not put secrets in instructions. Use connections and environment variables.

Exam traps

• Instructions are agent-wide. Prompts are individual reusable tools with inputs.
• Choosing a model in the test pane does not change production until you publish.
• Citations show sources. They do not make an answer more accurate.
• Foundry and Azure AI Search are different: one supplies models, the other indexed content.
• A longer instruction is not better. Conflicting rules cause inconsistent behaviour.

Practice questions

Choose an answer, then review the explanation to check your understanding.

1. A company requires that prompt-based responses come only from an approved model. Where do you pick that model?

1. A.In the channel settings
2. B.In the prompt model setting, from the available or Microsoft Foundry catalog
3. C.In the DLP policy
4. D.In the topic trigger

Show answer

Answer: B.A prompt lets you select the model it runs on, including models brought in from Microsoft Foundry.

2. The agent keeps answering questions outside its subject area. Which first fix is most direct?

1. A.Add a channel
2. B.Tighten the agent instructions to state scope and refusals
3. C.Delete knowledge sources
4. D.Switch to classic orchestration

Show answer

Answer: B.Instructions define scope, tone and what the agent must refuse.

3. You want the same summarising instruction reused in several agents and flows. What do you create?

1. A.A custom prompt tool with input variables
2. B.A new channel
3. C.A solution pipeline
4. D.A topic trigger phrase

Show answer

Answer: A.A prompt with input variables is a reusable AI tool that agents and flows can call.

4. How can users check where a generated answer came from?

1. A.Disable knowledge
2. B.Show citations in the response
3. C.Lower the model temperature
4. D.Use classic orchestration

Show answer

Answer: B.Citations link generated answers back to the knowledge source.