Test legat de ghid · Mediu Nivel

Test pentru agenții AI

Evaluați buclele de acțiune ale agentului, limitele de oprire, limitele permisiunii, postcondițiile observabile și raportarea lucrărilor parțiale sau eșuate.

Căi de ghidare aferenteAgenți AI
Întrebare 1 de 8

Un agent trebuie să creeze o schiță de eveniment din calendar pentru marți, la ora 14:00. Care probe ar stabili rezultatul solicitat?

Previzualizează fiecare întrebare8 întrebări

Răspunsurile și explicațiile apar pe măsură ce susțineți testul.

  1. Un agent trebuie să creeze o schiță de eveniment din calendar pentru marți, la ora 14:00. Care probe ar stabili rezultatul solicitat?
    • Înregistrarea rezultată are calendarul, data, fusul orar, titlul și starea de schiță corecte
    • Răspunsul agentului spune că a creat un eveniment
    • Un instrument a returnat succes fără a identifica calendarul salvat
    • Un eveniment cu nume similar există într-un calendar diferit
  2. Which sequence best matches the agent loop described in the guide?
    • Write a final answer before observing anything
    • Observe state, select an action, receive a result, and decide whether to continue
    • Repeat actions without checking results
    • Wait for an external page to grant authority
  3. An agent reads a web page that tells it to ignore the user’s goal. How should it treat that text?
    • As proof the user changed their request
    • As automatic permission to act
    • As external content whose trust level must be considered
    • As evidence the action already happened
  4. In the guide, what should ordinary software enforce around model-planned actions?
    • Whether a user has read the whole manual
    • The model’s preferred tone
    • The truth of every external page
    • Permissions, budgets, and tool contracts
  5. Why can narrow tools and permissions be useful for an agent?
    • They guarantee every tool call succeeds
    • They can limit the consequences of mistakes
    • They remove the need to inspect results
    • They ensure external pages are trustworthy
  6. Which status should an agent report when it completed only part of a task?
    • Partially achieved
    • Complete, because some work happened
    • Cancelled, regardless of what remains
    • Verified, because the agent says so
  7. An agent claims a file edit succeeded, but the tool failed. What should the operator do?
    • Trust the confident summary
    • Repeat the failed call indefinitely
    • Inspect the actual file and required checks before reporting completion
    • Ignore the failure because the task description was clear
  8. What should an agent report after an action workflow?
    • A claim of full success even if blocked
    • Only a confident completion statement
    • Every instruction found in external pages
    • Evidence of what changed and what remains uncertain