Plan and Configure Agent Solutions (Domain 1 — ~33%)
This domain tests your ability to make the right architectural choices before building and to correctly configure the core components of a Copilot Studio agent: knowledge sources, instructions, topics, entities, variables, and generative AI settings. Expect scenario-based questions that ask you to pick the right tool, the right knowledge source type, or the right variable scope for a given requirement.
Choosing the Right Platform: Copilot Studio vs. M365 Copilot Extensibility vs. Azure AI Foundry
This is a high-frequency decision question on the exam. Use this decision framework:
| Scenario | Choose |
|---|---|
| Build a standalone conversational agent with custom knowledge and topics | Copilot Studio |
| Extend Microsoft 365 Copilot (Teams, Outlook, Word) with plugins or declarative agents | M365 Copilot extensibility |
| Need full LLM control, custom model fine-tuning, or complex orchestration pipelines | Azure AI Foundry |
Trap: "Declarative agent" sounds like Copilot Studio, but declarative agents that surface inside Microsoft 365 Copilot chat belong to M365 extensibility. Copilot Studio agents can also be published to Teams, but the starting point is the agent, not the M365 Copilot surface.
Creating an Agent
Three creation paths, each with distinct exam implications:
- From template: Fastest path; template pre-populates topics, instructions, and sometimes knowledge. You still must configure environment and publish.
- From scratch: Full control; you define all topics and knowledge manually.
- From a URL or document: Copilot Studio crawls a public website or processes an uploaded file to seed the knowledge base. Crawled sites are limited to 2 levels deep by default; uploaded files support PDF, DOCX, and XLSX.
Key limit: A single agent supports up to 15 knowledge sources in the standard configuration. Enterprise connectors (ServiceNow, SAP) count separately under the enterprise knowledge category.
Configuring Generative AI Orchestration and Instructions
Generative orchestration lets the agent use an LLM to decide at runtime which topic or knowledge source to invoke, rather than relying solely on trigger-phrase matching. Enable it in Settings → Generative AI → Generative mode.
Instructions are a system-level prompt that shape the agent's persona, tone, and behavioral guardrails. Best practices for the exam:
- Instructions are processed as a system prompt, not a user turn — do not put dynamic runtime values here.
- Keep instructions focused: persona, scope constraints, and escalation behavior.
- Instructions interact with topics: if a topic handles a specific intent, the topic takes precedence over a purely instruction-driven response.
Prompt engineering fundamentals tested:
- Role definition, task framing, and output format constraints all belong in instructions.
- Avoid ambiguous negative constraints ("don't do bad things") — the exam favors specific, bounded instructions.
Adding and Managing Knowledge Sources
| Source Type | Category | Notes |
|---|---|---|
| SharePoint sites/libraries | Enterprise | Requires agent to be in same tenant; uses Microsoft Search index |
| Dataverse tables | Enterprise | Configured via table selection; respects row-level security |
| Public websites | Basic | Crawled at creation; refreshed periodically, not real-time |
| Uploaded files | Basic | Static snapshot; must re-upload to update content |
| ServiceNow / SAP | Enterprise (connector) | Requires premium licensing and connector configuration |
Sign in free to read the full guide
1 more section (about 636 words) continue below. The first domain's study guide is free for every AB-620 learner with a free account.