6.3. Configuring Agent Knowledge and Settings
💡 First Principle: An agent is only as smart as the knowledge you give it and as focused as the instructions you write. An agent with no configured knowledge is essentially a general Copilot chat — configuration is what makes it specialized.
Three configuration elements define how an agent behaves:
1. Knowledge Sources
Knowledge sources are the documents, sites, or data repositories the agent draws from when answering questions. Without explicit knowledge sources, the agent uses only its general AI training.
| Knowledge Source Type | Example | How to Configure |
|---|---|---|
| SharePoint site | Your HR policy intranet | Add the SharePoint URL to the agent's knowledge |
| Specific files | A product catalog PDF | Upload or link the file in the knowledge configuration |
| Website | Your public documentation site | Add the URL to the agent's knowledge sources |
| Graph data | Emails, meetings, calendar | Enable Microsoft Graph connector (where available) |
2. Instructions
Instructions are the system-level prompt that shapes the agent's overall behavior — its persona, its constraints, and its default approach. Think of instructions as the agent's standing rules.
Effective instruction examples:
- "You are an HR assistant for Contoso. Only answer questions about Contoso's official HR policies. If a question is outside HR scope, politely redirect to the relevant department."
- "Always respond in formal English. Keep answers concise — no more than 200 words. If you are unsure of an answer, say so and recommend the employee contact HR directly."
- "You help new employees navigate the onboarding process. Guide them step by step through tasks. Be encouraging and welcoming in tone."
3. Capabilities and Suggested Prompts
Capabilities control what actions the agent can take beyond text responses — for example, searching the web, accessing specific connectors, or triggering workflows.
Suggested prompts are pre-written questions that appear in the agent's interface to guide users toward common tasks. Well-chosen suggested prompts dramatically improve the user experience by showing people what the agent can do without requiring them to guess.
Example suggested prompts for an HR agent:
- "What is our company's parental leave policy?"
- "How do I submit an expense report?"
- "What are the steps to request PTO?"
⚠️ Exam Trap: Configuring an agent with knowledge sources does not mean the agent will answer exclusively from those sources unless you specify that in the instructions. If you want strict scoping ("only answer using this documentation"), you must write that into the instructions explicitly. Without such an instruction, the agent may blend its configured knowledge with general AI training.
Reflection Question: A compliance agent is configured with the company's regulatory policy documents as its knowledge source. Employees report that the agent sometimes gives answers that do not match the official policies. What configuration element is most likely missing or incomplete?
