AB-620 - Microsoft Certified AI Agent Builder Associate Cheat Sheet

Updated 2026-08-05

Next Topic: AI-103 - Microsoft Certified Azure AI Apps and Agents Developer Associate Cheat Sheet

🎯Take a practice test on this topic5 practice tests · 213 questions→

The AB-620 exam certifies your ability to design, build, and integrate enterprise-grade AI agents in Microsoft Copilot Studio, aimed at professional developers and ISV partners who already know Power Fx, Dataverse, and prompt engineering. Live in beta since April 2026 with general availability expected around July-August 2026, it tests three areas: Plan and configure agent solutions (30-35%), Integrate and extend agents in Copilot Studio (40-45%), and Test and manage agents (20-25%). The heaviest weight sits on integration: wiring agents to Microsoft Foundry, Model Context Protocol (MCP) servers, Fabric data agents, and orchestrating true multi-agent solutions over the Agent2Agent (A2A) protocol, not just building a single conversational flow. A passing score is 700 out of 1000. Many questions reward Microsoft's preferred pattern for a scenario, such as reaching for a Foundry agent when a task needs advanced reasoning or orchestration rather than rebuilding that logic inside a Copilot Studio topic.

What This Cheat Sheet Covers

This topic spans 15 focused tables and 175 indexed concepts, 5 practice tests with 213 questions. Below is a complete table-by-table outline of this topic, spanning foundational concepts through advanced details.

Table 1: Plan Enterprise Integration, Channels & Reusable ComponentsTable 2: Plan Identity, Responsible AI & GovernanceTable 3: Building Agent FlowsTable 4: Monitoring & Error Handling in Agent FlowsTable 5: Topic Building Blocks: Flows, Tools & VariablesTable 6: Advanced Agent Responses: Prompts, Knowledge, APIs & Generative AnswersTable 7: Response Formatting & Adaptive CardsTable 8: Connecting to Enterprise Knowledge SourcesTable 9: Computer-Use AgentsTable 10: MCP Tools, Custom Connectors & REST APIsTable 11: Multi-Agent Solution DesignTable 12: Integrating Foundry, Fabric & A2A Protocol AgentsTable 13: Integrating Agents with Azure: AI Search, Foundry Model Catalog & Application InsightsTable 14: Evaluating Agent PerformanceTable 15: Application Lifecycle Management (ALM) for Agents

Quick IndexSubscribe to unlock

A jump-to index of every table row in this cheat sheet.

Mind MapSubscribe to unlock

An interactive map of every table and concept in this topic.

Table 1: Plan Enterprise Integration, Channels & Reusable Components

Covers the AB-620 "Plan an agent solution" task: choosing an integration pattern for enterprise systems, picking channels and handling live-agent handoff plus channel security, packaging reusable components across agents, and designing for internal versus external audiences by matching channel and access posture to who the agent serves.

Concept	Example	Description
Power Platform Connectors	Prebuilt ServiceNow connector adds a "create ticket" action with zero dev time; a custom connector wraps an internal REST API once for every agent to reuse	• Prebuilt = configuration only; custom = build once, reuse across many agents. • Supports Power Fx and built-in parsing/error handling; not to be confused with a raw HTTP request, which can't be shared org-wide the same way.
HTTP Requests	`GET https://api.contoso.com/orders/{id}` with custom headers, when no prebuilt or custom connector exists yet	Fastest way to reach a service with no connector; the maker sets the URI, method, headers, and body on every request. Not reusable across agents like a custom connector.
Agent Flows	An agent flow validates a support ticket, calls a low-code connector action, then pauses for a human-in-the-loop approval	Deterministic, low-code chains of connector actions built for high throughput; can include human-in-the-loop steps. Billed in Copilot credits, unlike Power Automate cloud flows' per-user or per-process licensing.
Bot Framework Skills	A pro-code C# skill registered in Copilot Studio, then triggered as a node inside a topic	Reusable, pro-code conversational building blocks that run synchronously in Azure AI Bot Service. Requires an Azure subscription and its own Application Lifecycle Management outside Power Platform.
Integration Planning Worksheet	ServiceNow row: Called by "Service Desk KB topic", Called with "Workflow", expected daily volume 5,000, expected peak 1,000	For every planned integration, document the connected system, what calls it, how it's called, and expected daily/peak volume, not just the connection type, so throughput needs surface before build.
Channels and Clients	Teams and Microsoft 365 Copilot ship a built-in client; a custom web app needs the Direct Line API plus a client like React Web Chat	A channel is the integration point where a user reaches the agent; native channels already include a client, but web/custom-app channels need one selected. Markdown and Adaptive Card support vary by client and channel.
Live Agent Handoff Patterns	Bot-as-an-Agent puts the Engagement Hub Chat Canvas in front; Bot-in-the-Loop keeps the Copilot Studio Chat Canvas in front and routes through a Bot Framework skill	• Bot-as-an-Agent: easiest setup, but no whisper mode unless the hub supports it. • Bot-in-the-Loop: full control of bot messages plus whisper mode (agent assist), but heavy pro-dev customization.
Web Channel Security	Settings > Security > Web channel security > turn on Require secured access, then exchange the secret for a token before any browser-side code	Demo and Custom website channels are open by default; anyone who knows the agent ID can chat. Use a secret for service-to-service calls, but exchange it for a token wherever code is visible to a browser or mobile app.
Component Collections	An HR topic lives in the "Corporate Services" collection; editing it once updates every agent connected to that collection	Reusable bundles of topics, knowledge, tools, and other components shared across multiple agents and environments, not scoped to one agent. Only the creator or a system customizer/admin can edit a shared component.
Primary Agent	Disconnect every agent but one from a collection, then Set primary agent to lock that collection to just that agent	Restricts a component collection to one owning agent, protecting sensitive or large collections from unintended reuse. Only settable once a single agent remains connected.
Solutions for Component Collections	A collection that includes a custom connector or environment variables gets packaged into a Solution, then exported as Managed or Unmanaged	A direct collection export always becomes managed; wrap it in a Power Platform Solution instead when it carries connectors or environment variables, or when an unmanaged package is needed.
Agent Library	Add the prebuilt Document Extraction component (an AI Builder prompt) to any agent as a managed, read-only solution	Prebuilt, Microsoft-authored component collections (Document Extraction, ServiceNow Ticket Assistant, and more). Behavior is shaped through agent instructions and tool input defaults, not by editing the underlying prompt or flow.
Audience-Driven Channel Selection	Support today happens over Outlook, not chat, so the agent is designed around complete email-style replies instead of a chat-first topic	Pick a channel based on where the audience already works, not on which is fastest to deploy. A channel also sets user expectations around latency and turn-taking, and forcing a chat-first design onto an email-driven audience hurts adoption.
Internal vs External Audience Design	A public FAQ website for anonymous visitors uses No authentication for reach; an internal HR agent for employees requires authentication	No authentication maximizes accessibility for broad, anonymous external audiences but drops user-identity context, so any action needing auth becomes unavailable. Internal agents or agents for specific users should apply authentication and governance controls instead.

Table 2: Plan Identity, Responsible AI & Governance

Plan and configure agent solutions (30-35%) > Plan an agent solution: the identity, responsible AI, and governance slice, covering how end users and agent tools authenticate, Microsoft's responsible AI practices, and the data policies and access controls administrators layer around an agent. Enterprise integration, channels, and reusable components are covered in the sibling table for this same task.

Concept	Example	Description
End-User Authentication Options	`Settings > Security > Authentication`: No authentication, Authenticate with Microsoft, or Authenticate manually	Configured per agent. No authentication lets anyone with the link chat with it; Authenticate with Microsoft auto-configures Entra ID but only works on the Teams + Microsoft 365 channel; Authenticate manually supports Entra ID or any OAuth 2.0 provider on any channel.

The rest of this section is part of the full cheat sheet

Table 3: Building Agent Flows

AB-620 task area Create and monitor agent flows in Copilot Studio: how a flow is created, wired as an agent tool, given actions and connectors, and passed input/output parameters. Monitoring and error handling for these flows are covered in Table 4.

Concept	Example	Description
Agent Flow	`Trigger: new email arrives -> Action: post message in Teams`	Deterministic, rule-based automation: the same input always produces the same output. •Not to be confused with generative orchestration, whose LLM planner can vary its plan.

The rest of this section is part of the full cheat sheet

Table 4: Monitoring & Error Handling in Agent Flows

AB-620 domain "Plan and configure agent solutions": tracking an agent flow's run history and performance, and building error handling into the flow itself. Covers the Activity/Analytics tabs and owner failure alerts for monitoring, plus Run After branching, Try/Catch scopes, retry policies, the Request for Information action, and the 100-second synchronous timeout for error handling.

Concept	Example	Description
Activity Tab (Flow Run History)	Activity tab, select a failed run, step through actions, Resubmit or Cancel	Full run by run history of an agent flow: status, duration, and the ability to resubmit or cancel a run. Not to be confused with the Overview tab, which shows only the most recent runs.

The rest of this section is part of the full cheat sheet

Table 1: Plan Enterprise Integration, Channels & Reusable Components

Concept	Example	Description
Power Platform Connectors	Prebuilt ServiceNow connector adds a "create ticket" action with zero dev time; a custom connector wraps an internal REST API once for every agent to reuse	• Prebuilt = configuration only; custom = build once, reuse across many agents. • Supports Power Fx and built-in parsing/error handling; not to be confused with a raw HTTP request, which can't be shared org-wide the same way.
HTTP Requests	`GET https://api.contoso.com/orders/{id}` with custom headers, when no prebuilt or custom connector exists yet	Fastest way to reach a service with no connector; the maker sets the URI, method, headers, and body on every request. Not reusable across agents like a custom connector.
Agent Flows	An agent flow validates a support ticket, calls a low-code connector action, then pauses for a human-in-the-loop approval	Deterministic, low-code chains of connector actions built for high throughput; can include human-in-the-loop steps. Billed in Copilot credits, unlike Power Automate cloud flows' per-user or per-process licensing.
Bot Framework Skills	A pro-code C# skill registered in Copilot Studio, then triggered as a node inside a topic	Reusable, pro-code conversational building blocks that run synchronously in Azure AI Bot Service. Requires an Azure subscription and its own Application Lifecycle Management outside Power Platform.
Integration Planning Worksheet	ServiceNow row: Called by "Service Desk KB topic", Called with "Workflow", expected daily volume 5,000, expected peak 1,000	For every planned integration, document the connected system, what calls it, how it's called, and expected daily/peak volume, not just the connection type, so throughput needs surface before build.
Channels and Clients	Teams and Microsoft 365 Copilot ship a built-in client; a custom web app needs the Direct Line API plus a client like React Web Chat	A channel is the integration point where a user reaches the agent; native channels already include a client, but web/custom-app channels need one selected. Markdown and Adaptive Card support vary by client and channel.
Live Agent Handoff Patterns	Bot-as-an-Agent puts the Engagement Hub Chat Canvas in front; Bot-in-the-Loop keeps the Copilot Studio Chat Canvas in front and routes through a Bot Framework skill	• Bot-as-an-Agent: easiest setup, but no whisper mode unless the hub supports it. • Bot-in-the-Loop: full control of bot messages plus whisper mode (agent assist), but heavy pro-dev customization.
Web Channel Security	Settings > Security > Web channel security > turn on Require secured access, then exchange the secret for a token before any browser-side code	Demo and Custom website channels are open by default; anyone who knows the agent ID can chat. Use a secret for service-to-service calls, but exchange it for a token wherever code is visible to a browser or mobile app.
Component Collections	An HR topic lives in the "Corporate Services" collection; editing it once updates every agent connected to that collection	Reusable bundles of topics, knowledge, tools, and other components shared across multiple agents and environments, not scoped to one agent. Only the creator or a system customizer/admin can edit a shared component.
Primary Agent	Disconnect every agent but one from a collection, then Set primary agent to lock that collection to just that agent	Restricts a component collection to one owning agent, protecting sensitive or large collections from unintended reuse. Only settable once a single agent remains connected.
Solutions for Component Collections	A collection that includes a custom connector or environment variables gets packaged into a Solution, then exported as Managed or Unmanaged	A direct collection export always becomes managed; wrap it in a Power Platform Solution instead when it carries connectors or environment variables, or when an unmanaged package is needed.
Agent Library	Add the prebuilt Document Extraction component (an AI Builder prompt) to any agent as a managed, read-only solution	Prebuilt, Microsoft-authored component collections (Document Extraction, ServiceNow Ticket Assistant, and more). Behavior is shaped through agent instructions and tool input defaults, not by editing the underlying prompt or flow.
Audience-Driven Channel Selection	Support today happens over Outlook, not chat, so the agent is designed around complete email-style replies instead of a chat-first topic	Pick a channel based on where the audience already works, not on which is fastest to deploy. A channel also sets user expectations around latency and turn-taking, and forcing a chat-first design onto an email-driven audience hurts adoption.
Internal vs External Audience Design	A public FAQ website for anonymous visitors uses No authentication for reach; an internal HR agent for employees requires authentication	No authentication maximizes accessibility for broad, anonymous external audiences but drops user-identity context, so any action needing auth becomes unavailable. Internal agents or agents for specific users should apply authentication and governance controls instead.

Table 2: Plan Identity, Responsible AI & Governance

Concept	Example	Description
End-User Authentication Options	`Settings > Security > Authentication`: No authentication, Authenticate with Microsoft, or Authenticate manually	Configured per agent. No authentication lets anyone with the link chat with it; Authenticate with Microsoft auto-configures Entra ID but only works on the Teams + Microsoft 365 channel; Authenticate manually supports Entra ID or any OAuth 2.0 provider on any channel.

Table 3: Building Agent Flows

Concept	Example	Description
Agent Flow	`Trigger: new email arrives -> Action: post message in Teams`	Deterministic, rule-based automation: the same input always produces the same output. •Not to be confused with generative orchestration, whose LLM planner can vary its plan.

Table 4: Monitoring & Error Handling in Agent Flows

Concept	Example	Description
Activity Tab (Flow Run History)	Activity tab, select a failed run, step through actions, Resubmit or Cancel	Full run by run history of an agent flow: status, duration, and the ability to resubmit or cancel a run. Not to be confused with the Overview tab, which shows only the most recent runs.