Free AI-500 practice test: 30 questions with answers

Scheduled evaluations on a fixed test set detect quality changes early, and threshold alerts turn the signal into action. Waiting for complaints, disabling evaluation and retraining without tests do not detect regression reliably.

Section-aware chunks with overlap and metadata let retrieval match the precise clause and filter by document attributes. Whole-document chunks dilute relevance, sentence-level chunks lose context, and removing metadata reduces filtering precision.

Structured facts with validation against sources prevent each rewrite from drifting, because the stored records remain the reference. Relying on recall, summarizing only a summary, and raising temperature all increase drift.

Schema validation catches malformed results, and retry or fallback policies keep the workflow safe. Forwarding raw data propagates errors, disabling the tool removes a needed capability, and guessing produces unreliable outputs.

API Management can front MCP servers implemented on Azure Functions or Logic Apps, adding authentication, throttling and monitoring. Static storage has no compute, and an unauthenticated laptop script is neither managed nor secure.

Partitioning by task type and running independent work in parallel removes the bottleneck while rate limits protect shared services. Removing coordination loses control, duplicate orchestrators create conflicts, and a single serial worker keeps the bottleneck.

A2A is designed for agents to discover and communicate with each other across vendors, and it can be secured with standard authentication. The other protocols are not agent interoperability standards.

Iteration and tool-call caps stop unbounded loops, and batching with rate limits reduces repeated work and protects shared services. Unlimited agent creation, indefinite retries and removed timeouts all create runaway cost or stalls.

The on-behalf-of flow exchanges the user's token for a token scoped to the downstream API, so the user's password is never stored. Shared passwords, URL-embedded keys and shared administrator accounts each break these requirements.

Key Vault with managed identity provides centralized secrets, scheduled rotation and access control without redeploying images. Embedded, shared or prompt-based secrets cannot be rotated safely.

A multi-point guardrail strategy checks tool responses and final outputs, so personal data is caught on both paths before it reaches the user. Checking only the question, trusting the model and auditing monthly leave the data exposed.

A canary release sends a small share of traffic to the new version and increases it as metrics stay healthy, with rollback if they degrade. A big-bang release gives no early warning, and manual or unmonitored releases remove the safeguard.

Data residency and audit logs satisfy location and accountability requirements, and customer-managed keys give control over encryption at rest. Personal folders, public logs and short audit retention all fail common regulatory expectations.

The AI Red Teaming Agent runs automated adversarial tests that probe for unsafe behaviour before release, which is shift-left security. Billing charts, calendars and printed policies do not test the agent.

A custom guardrail encodes a domain-specific rule such as no internal pricing and applies it to outputs. General filters target other categories, temperature changes randomness, and asking politely in the prompt is not enforcement.

DTAP moves releases through development, test, acceptance and production, so each stage catches different problems before users are affected. Skipping stages or going straight to production removes those checks.

A judge model scores answers against a rubric, and human spot checks calibrate the judge. Character counts, single-word checks and CPU metrics do not measure answer quality.

Per-request token limits and loop controls stop runaway generation and repeated tool calls that inflate usage. Hiding traces, enlarging context windows and asking for repeated reasoning all increase token consumption.

A shared correlation ID across agents and tools lets operators follow one request through the workflow and find the failing step. Unlinked logs, blind restarts and sampling only successes hide the failing component.

Tagging usage by team and workload allows accurate cost reports and chargebacks based on consumption. Shared keys and equal splits or estimates do not reflect real usage.

Entity disappearance across turns and restatement accuracy directly show whether the agent keeps track of who or what is being discussed. More tools, emoji counts and unmanaged verbatim history do not measure entity continuity.

Testing quality and latency together against service-level targets on representative traffic shows whether the trade-off is acceptable. Ignoring latency, laptop-only measurements and commit counts do not inform the decision.

Bounded parallelism with exponential backoff raises throughput while respecting rate limits. Unbounded bursts cause more throttling, serial execution is slow, and dropping failed tasks loses work.

Sequential orchestration passes each approved output to the next agent, which matches dependent, non-concurrent steps. Parallel fan-out would run dependent steps before inputs exist, peer-to-peer broadcast gives no fixed order, and hub-and-spoke without shared state loses the handoff context.

A threshold-based approval gate keeps low-risk refunds fast while routing high-value actions to a person, and logging provides accountability. Unlimited autonomy and removed logging each remove control over consequential actions.

Distinct identities with scoped role assignments give least privilege and per-agent audit trails. Shared principals and personal credentials blur accountability, and prompt-only limits are not an access control.

Multi-tier memory separates session, team and long-term data with suitable lifecycle and retention controls. A single ever-growing prompt is costly and unbounded, a shared text file has no isolation, and retraining for each memory update is impractical.

Zero Trust with per-agent identity and segmented paths limits what a compromised agent can reach, which prevents lateral movement. Flat trust, shared secrets and disabled authentication each make lateral movement easier.

Replay capture stores the reasoning path, tool calls and their inputs and outputs so the failure can be reproduced. Averages and counts do not show what happened, and the final answer omits the intermediate steps.

A tool registry with per-agent scopes controls what each agent may call, and an integration layer routes calls and records results for governance. A shared unrestricted credential, random selection and unlogged chat remove the controls production systems require.