Free AB-620 practice test: 30 questions with answers

Delegated authentication keeps the API's authorization tied to each user, which prevents privilege escalation through a shared account. A shared admin account, anonymous access, and stored passwords each break accountability and least privilege.

Planning channels and their authentication and display needs lets one agent serve several channels consistently. Avatar color is cosmetic, and copying separate agents creates maintenance drift.

Retrieval must respect the user's permissions, and sensitivity labels and source permissions shape what can be surfaced. Public publication, disabled logging, and tenant-wide sharing each expose confidential content or remove oversight.

Reusable components centralize shared logic so fixes apply everywhere it is used. Copy-paste duplication causes drift, removing error handling reduces reliability, and personal documents do not run.

A responsible AI assessment examines risks to people and how oversight and transparency are handled before release. Icon size, speed alone, and skipping review do not address that risk.

A representative test set with expected outcomes is the basis for measurable evaluation. Slides, personal prompts, and production database copies do not define expected behavior.

Grounding failures usually come from retrieval quality or instructions that allow unsupported claims, so these are the first things to review. Report colors, case counts alone, and author names do not explain the failures.

Solutions package the components so they can be moved and versioned consistently through environments, often with pipelines. Manual recreation, icon exports, and undocumented edits create drift and risk.

A versioned test set rerun after changes detects regressions, and comparing against acceptance criteria makes results meaningful. Adjusting expectations to pass tests, testing one question, and deleting failures all hide problems.

Subjective quality needs human judgment against a rubric, and automated checks can cover objective criteria. Word counts, speed, and absence of complaints do not measure helpfulness.

Copilot connectors make indexed enterprise content searchable with permissions preserved, which fits the requirement. Manual slides, hard-coded answers, and scrapers do not provide governed enterprise retrieval.

Configuring the MCP server as a tool source lets the agent discover and call the server's tools. Separate topics per method, copied code, and unauthenticated exposure do not meet the requirement securely.

Azure AI Search provides governed, indexed retrieval that generative answers can use. Spreadsheets read by email, unindexed crawls, and hard-coded answers do not meet the governed retrieval need.

REST APIs and custom connectors give the agent defined, governed actions against external services. Copying source code, hard-coded responses, and manual email do not provide a callable integration.

Integrating the existing Fabric data agent reuses tested logic and keeps a single source of analytics behavior. Exporting static files, rebuilding logic, and disabling the agent duplicate or discard working capability.

Adaptive cards define structured layouts and actions that render consistently across supported channels. Plain text, duplicate agents, and embedded HTML do not provide a portable, structured card format.

The generative answers node uses configured knowledge sources to produce grounded responses with references. A fixed message, a shoe-size check, and a variable store do not generate grounded answers.

A conversation variable keeps the value across the steps of one conversation, and a flow-level global variable can also be read later. Hard-coding repeated values is fragile, passwords in messages are unsafe, and separate agents per step are unnecessary.

Custom prompts can use a registered catalog model, which keeps the model choice governed and traceable. Copying weights into text, client-side model choice, and unregistered models do not work within the agent platform.

The Send HTTP request node makes the call and lets the builder map response values into variables for display. Typing an address in a message does not call the service, and condition or name-only variables do not invoke it.

A human-in-the-loop flow pauses for an explicit decision before the action runs. Automatic approval, self-approval, and monthly schedules do not provide the required human checkpoint.

Error handling with a fallback path informs the user and records the failure for diagnosis. Silent failure, endless retries, and fake success messages hide problems and mislead users.

Run history and failure details show frequency and failure points accurately. Calendars, guesses, and paper logs do not reflect actual flow runs.

Typed input and output parameters define the contract between the topic and the flow, which keeps values reliable. Untyped globals, comments, and screenshots do not define that contract.

Connection references and an authentication model matched to the run context keep credentials managed and environment-specific. Descriptions are not secure storage, a personal connection breaks accountability, and disabling authentication exposes the data.

Application Insights collects telemetry such as latency and exceptions in a standard monitoring system. Printed reports, typed spreadsheets, and per-message emails are slow and incomplete.

Correlation identifiers link every step of one conversation across logs, which makes intermittent failures traceable. Daily counts, short retention, and browser restarts do not reveal the path of a failing conversation.

Tagging events with topic and time lets the dashboard group failures by topic and day. A monthly total, weekly screenshots, and deleted logs cannot produce that breakdown.

A metric alert rule with a threshold and time window notifies the team automatically when the condition holds. Manual checks, comments, and disabling measurement do not provide timely alerting.

Per-action logs with timestamps and identity show what was done and by which agent. Memory summaries and a final screenshot miss intermediate actions.