Test and Manage Agents (27%)
This domain tests your ability to move an agent from development to production safely. Expect questions on validating agent behavior before release, controlling who can build and use agents, enforcing data policies, tracking quality through analytics, and keeping agents compliant over time. The emphasis is operational: knowing which tool does what and at what stage of the lifecycle.
Testing Agents
The test chat pane in Copilot Studio is the first-line tool. It runs inside the Studio canvas and lets you exercise topics interactively without publishing. Key capability: the Reset button clears session state so you can re-test trigger phrases from a clean context.
Topic/conversation path debugging uses the conversation map view—each node highlights as it executes, letting you see exactly which branch was taken. Watch for trap questions: the test pane reflects saved content, not unsaved drafts; always save before testing.
Tracing goes deeper. The trace panel exposes Power Automate flow run IDs for any action called during a turn, so you can cross-reference run history in Power Automate. For generative AI answers, tracing shows which knowledge source was cited. Exam distinction: tracing is diagnostic, not a live monitoring tool—it covers the current session only.
Publishing Agents to Channels
Publishing is a deliberate, manual step—saving a topic does not push changes to end users. You must click Publish to promote the current version.
| Channel | Key consideration |
|---|---|
| Microsoft Teams | Requires Teams app submission; admin approval may be needed for org-wide deployment |
| Website / custom app | Generates an embed code (iframe or Direct Line token); token scope controls which bot is accessed |
| Mobile apps | Use the Direct Line API; token-based auth, refresh interval matters for session continuity |
Trap: A published agent on Teams and a published agent on a website share the same underlying bot but can have channel-specific configurations (e.g., authentication enabled for Teams, anonymous for web). Changing a channel setting does not require a full republish of all channels.
Security Roles in Copilot Studio
Three roles govern what users can do in a Copilot Studio environment:
- System Administrator / Environment Admin: Full control—create, delete, publish agents, manage connectors, assign roles.
- Maker (Environment Maker): Can create and edit agents they own; cannot manage environment-level settings or publish to org-wide channels without additional configuration.
- Agent Contributor: Can edit an agent they've been shared with but cannot publish it or manage security settings. This is the minimum role for collaborative development on a specific agent.
Exam trap: "Agent Contributor" is agent-scoped, not environment-scoped. A user can be an Agent Contributor on one agent and have no access to others in the same environment.
Data Loss Prevention (DLP)
DLP policies in the Power Platform admin center classify connectors into Business, Non-Business, and Blocked buckets. Connectors in different buckets cannot be used together in the same agent or flow.
Critical exam points:
- Blocking a connector that an existing agent already uses will break that agent at runtime—not at design time. The agent saves but fails when the blocked action executes.
- Generative AI features (like knowledge sources using web search) may rely on connectors that must be explicitly allowed.
- DLP policies apply at the environment or tenant level; tenant-level policies override environment-level ones.
Unlock with a paid plan
1 more section (about 697 words) continue below. Every paid plan unlocks all AB-620 study guides, the full question bank and mock exams, from $4.97/month.