Free AB-620 Exam Flashcards

Authenticate manually. Authenticate with Microsoft sets up Entra ID automatically but is built for the Teams + Microsoft 365 channel, and Microsoft says to choose manual for other channels. Manual mode supports Microsoft Entra ID V2 and generic OAuth 2 providers and exposes User.AccessToken and User.IsLoggedIn.

The topic-level setting on the generative answers node wins at runtime. With nothing set there, the node inherits the agent's Generative AI setting, whose default is High. A prompt tool carries its own moderation level, defaulting to Moderate.

Block Chat without Microsoft Entra ID authentication in Copilot Studio. Makers are then limited to Authenticate with Microsoft or Authenticate manually, and the No authentication option disappears from the agent's Security settings.

Direct Line channels in Copilot Studio. That one connector covers the demo website, custom websites, mobile apps and other Direct Line channels. Teams is governed separately by the Microsoft Teams + Microsoft 365 Channel connector.

When it needs its own publishing, channels, model settings or ALM process, or must be reused by several main agents. Also split when the main agent passes roughly 30 to 40 choices across tools, topics and agents and selection accuracy degrades.

Agent flows are deterministic. The same input follows the same rule-based path and produces the same output, which suits repeatable processes such as ticket creation or approvals where predictability matters more than the orchestrator's flexibility.

The When an agent calls the flow trigger and a Respond to the agent action. The flow must also be published, respond synchronously with Asynchronous response turned off, and finish within the 100-second action limit. Other triggers run instantly, on a schedule or on events.

Human in the loop actions, such as approval requests. They insert a deterministic checkpoint before a consequential step. The other categories are AI capabilities, built-in tools for looping and branching, and connectors.

A topic call consumes one Classic answer plus the flow's actions. A generative orchestration call consumes one Autonomous action plus the actions. Runs started from the flow designer or the agent's test chat consume no Copilot Studio capacity.

Power Platform admin center, under Licensing then Copilot Studio, lists Agent flow actions per flow. Once prepaid Copilot Studio capacity is fully consumed, new runs are blocked until capacity frees up or pay-as-you-go billing is enabled; in-flight runs still complete.

It runs a knowledge search at a chosen point inside an authored topic, so grounding happens for a specific intent instead of only as a fallback. Every new agent already has one inside the Conversational boosting system topic.

Sources configured on the generative answers node take priority; agent-level sources act as a fallback. Note that with generative orchestration the agent bypasses the Conversational boosting topic entirely, so customizations made there are ignored.

The agent chooses. Selection then rests on the topic's description, name and parameter names rather than trigger phrases. Copilot Studio auto-generates descriptions from existing trigger phrases, but rewrite them so similar topics stay distinguishable.

Add a Parse value node, choose From Sample Data, and paste sample JSON to derive the schema. The output variable becomes a Record you can address with Power Fx. Parse value also types runtime objects such as System.Activity.ChannelData.

From a Message or Question node select Add, then Adaptive card, then Edit adaptive card to open the Adaptive Card designer. Build from elements or paste the payload in the card payload editor. Two or more cards render as a carousel by default.

In the second topic mark the variable Receive values from other topics, then pass a value on the Redirect node. Return values to original topics sends an answer back. Converting a topic variable to a global variable is one-way.

Temperature, a slider from 0 to 1 that defaults to 0 for consistent output; higher values give more varied text. The slider is disabled for GPT-5 reasoning, and the content moderation slider is unavailable for Anthropic and Azure AI Foundry models.

SharePoint, Dataverse and enterprise data through Copilot connectors all query using the agent user's Microsoft Entra ID, so each person sees only content they can access. Public website and uploaded document sources carry no per-user authentication.

Classic orchestration caps a node at four SharePoint URLs, four website URLs and two Dataverse sources. Generative mode allows 25 SharePoint URLs, 25 websites and unlimited Dataverse or connector sources, and an internal model filters sources once more than 25 exist; uploaded files are excluded from that count.

Any turn in which the agent used no knowledge source and no tool is blocked and the fallback topic triggers. The agent also withholds answers that carry no in-text citation, so follow-ups answered from conversation history alone can fail intermittently.

A vector index built with integrated vectorization, optionally with the semantic ranker configured. Add it through a formal data connection under Add knowledge rather than a hand-entered endpoint and key, and attach only one vector index per connection.

A public website source restricts Bing to the sites you listed. Web Search, also shown as Use information from the web, searches everything Bing indexes and interleaves those results with your listed sites. It needs generative orchestration and uses Grounding with Bing Search.

When the system exposes no API and only a human-usable interface. Computer use drives a Windows machine with a virtual mouse and keyboard from natural-language instructions, and its vision model adapts when the interface changes. It requires generative orchestration.

It blocks interaction with sites and apps that are not listed, but not opening them. The model can still navigate to an unlisted page, and every attempt to act there fails. Enforce HTTPS separately stops it working with plain HTTP sites.

Only the Streamable HTTP transport. Server-sent events were deprecated in the MCP specification and Copilot Studio dropped SSE support after August 2025. A custom-connector OpenAPI file declares the transport with x-ms-agentic-protocol: mcp-streamable-1.0.

None, API key, or OAuth 2.0. An API key travels in a named header or query parameter. OAuth 2.0 offers dynamic discovery, dynamic client registration with endpoints you supply, or manual setup with client ID, secret, authorization, token and refresh URLs.

MCP servers reach Copilot Studio over Power Platform connector infrastructure, so a data policy that regulates connectors also regulates the MCP server and its tools. Enforcement is real time: the maker sees the violation banner and the Publish button becomes unavailable until the policy or the agent changes.

Its name and description, plus the names and descriptions of its input and output parameters. Write one or two active-voice sentences saying what the tool does and what it cannot do, and avoid names that overlap with an existing tool.

Child agents created inside the agent, other Copilot Studio agents in the same environment, and external agents over the A2A protocol. Microsoft Foundry agents, Fabric data agents and Microsoft 365 Agents SDK agents can also be connected, and those three are in public preview.

A connected agent. Child agents group tools, instructions and knowledge inside one agent with no separate settings or deployment. Connected agents get their own configuration, model, channels and lifecycle, and can be reused by more than one main agent.

Latency from extra orchestration hops. The main agent plans, selects a connected agent, and that agent then runs its own orchestration over its own tools. Splitting also enlarges the testing, management and governance surface.

No. An agent acting as a main agent with one or more connected agents cannot itself be a connected agent for a second main agent. Any agent with no connected agents of its own can be reused by multiple main agents.

The form needs the agent's message endpoint URL, not the card URL. When a valid card sits at the standard .well-known path, Copilot Studio pulls the name and description automatically; otherwise enter both by hand so the orchestrator knows when to delegate.

A2A is designed for agent workflows: multiturn interactions, rich contextual metadata such as a contextId and chat history, and interoperability across frameworks. An HTTP connector calls an API with limited context. Use MCP instead when you need tools and resources.

Fabric data agents. They cannot be redirected to from within a topic, cannot be referenced with a slash in the main agent's instructions, and do not work when the main agent is deployed to Microsoft 365 Copilot.

Embed it in a generative answers node and choose the Classic data option, then add the Azure OpenAI connection. Generative orchestration does not support custom data or Bing Custom Search as agent-level knowledge, so those sources work only inside a node.

Access to external models for the environment. Anthropic and xAI models are hosted outside Microsoft under their own terms, and the prompt builder's content moderation slider is unavailable for them. In the prompt builder, Claude models still count as experimental even though they display no tag.

Settings, then Advanced, then paste the Application Insights connection string. Enable logging captures incoming and outgoing messages and events, Log conversation details adds user ID, name and message text, and Node execution events logs each node run.

Filter on the designMode custom dimension that every event carries and keep only rows where it equals False. Telemetry is logged for all conversations, including testing inside Copilot Studio. A prebuilt Copilot Studio Dashboard workbook ships in the Workbooks gallery.

An evaluation runs a saved test set of single-response or conversation cases against defined methods and returns repeatable, comparable results. The test panel is manual and one scenario at a time, and its activity is excluded from the Monitor page.

General quality. It scores relevance, groundedness, completeness and abstention, and a response must satisfy all four criteria to be rated high quality. It needs no expected answers, and every test set starts with this method by default.

Use the Tool use method. It passes only when the agent invoked the expected tools or topics while generating the answer, and fails otherwise. A test case with no expected tools returns an Invalid result rather than a pass or fail.

Compare meaning, which judges intent similarity against your expected answer and defaults to a pass score of 50. Text similarity is the wrong choice here: it uses cosine similarity on wording and sentence structure, so use it only when phrasing itself matters.

Tool use, keyword match, exact match and custom return pass or fail. General quality, compare meaning and text similarity return a score, and you set the pass threshold for compare meaning and text similarity. Keyword match can require any or all keywords.

You cannot export it. Managed solutions cannot be exported, so you must create a new unmanaged solution and add the agent to it again. Solutions are unmanaged when created, and converting one to managed is not reversible.

Open the solution, select the agent's three dots, choose Advanced, then Add required objects to pull dependencies such as topics, flows and environment variables into the solution. Never delete or edit an agent's components directly from the solution.

Import the custom connector first, then the connection reference along with the agent solution. After importing, reconfigure user authentication for the agent and publish it before sharing. The agent icon can take up to 24 hours to appear everywhere.

Use an environment variable that references an Azure Key Vault secret. Grant the Key Vault Secrets User role to the Microsoft Copilot Studio Service application and tag the secret with AllowedEnvironments or AllowedAgents. Anyone who can edit the agent could still surface the value.

The solution is exported once, when the maker submits the deployment, and that same tamper-protected artifact moves through every stage in order. The system never re-exports for a later stage, so no customization can bypass a QA or approval stage.

Connections, connection references and environment variables, all validated before deployment starts so no manual post-processing is needed. Pipelines deploy managed solutions only, do not move Dataverse table data, and cannot deploy across tenants.