Quizzes

Tools & Function Calling quiz

24 questions from the Tools & Function Calling track. Score 80% or higher to unlock a certificate.

1. Why Tools

   1. What do tools add that model weights cannot?

   • Extra adjectives in the system prompt
   • Fresh data, side effects, and computations your runtime actually runs
   • A guarantee the model never errs
   • Unlimited context without tokens

2. Model vs Runtime

   2. Who decides that a refund really happened?

   • The assistant’s last sentence
   • Your runtime log after the refund function returned
   • The system prompt
   • A confident thought

3. Read vs Write Tools

   3. Which tools are usually safe to retry without extra keys?

   • Refund and send_email
   • Read-only getters that do not change the world
   • Any tool with a friendly name
   • Tools the model called twice

4. What Is Not a Tool

   4. When is something a real tool?

   • When the system prompt names it
   • When your runtime can validate args and execute it
   • When the model writes “I called it”
   • When the UI has a sparkle icon

5. JSON Schema for Tools

   5. When should you execute a tool call?

   • As soon as the model emits a name
   • After arguments pass your schema validator
   • After printing the system prompt
   • Only if the thought looks confident

6. Required Fields and Extra Keys

   6. Why reject unknown fields?

   • To make the JSON prettier
   • So the model or an injected doc cannot smuggle extra arguments
   • Because enums are optional
   • Only when the thought is long

7. Descriptions Are Prompts

   7. What belongs in a tool description?

   • Only the function name in all caps
   • When to use it, when not to, and which ids it accepts
   • A 40-page SDK
   • A joke

8. Function Calling

   8. Who executes a function call?

   • The GPU, automatically
   • Your runtime, after parsing name plus args and checking a registry
   • The user, by reading the thought
   • The vector database

9. The Dispatcher

   9. How should an agent look up a tool?

   • Run the name string as Python
   • Load attributes off the os module
   • A dict (or table) you registered by hand
   • Whatever the last assistant sentence mentioned

10. Coerce and Structured Errors

    10. What should a tool return when job_id is -1?

    • A Python traceback
    • A small JSON error with a code, the field, and a hint
    • An empty string
    • “Sure, job -1 is fine”

11. Parallel Tool Calls

    11. When is it safe to run two tool calls at the same time?

    • Always, because GPUs are fast
    • When they are independent reads (or writes that are truly idempotent)
    • When both names start with get
    • When the model asked nicely

12. Observations and Truncation

    12. What should you do with a huge tool result?

    • Paste all of it into the next prompt
    • Truncate, set truncated true, and offer a way to fetch more
    • Drop it and hope
    • Put it in the system prompt

13. Designing Tools

    13. Which tool shape should you ship?

    • Small, typed functions with structured errors
    • One run(command: str) that the model should “use carefully”
    • Tools named after your feelings
    • Tools that never return errors, only poetry

14. Idempotency

    14. What should a second refund of the same invoice do?

    • Charge twice
    • Return the same receipt and mark the call idempotent
    • Crash
    • Ask the model to think harder

15. Timeouts, Retries, and Size Caps

    15. After a refund times out, what is the safe next step?

    • Call refund again with a new id
    • Poll status or retry with the same idempotency key
    • Run it 20 more times
    • Ignore the error

16. Code Interpreters

    16. How should an agent run model-written Python?

    • In the same process that holds production keys
    • In a sandbox: allowlisted ops or a locked-down runtime, no network, limits, logged code
    • By pasting it into production bash
    • Only if the thought said please

17. Browser and Computer Use

    17. What is a browser/computer-use agent, mechanically?

    • A new kind of GPU
    • A loop over GUI state with discrete action tools and a success predicate
    • A screenshot pasted into Slack
    • An agent with no stop condition

18. Model Context Protocol (MCP)

    18. What is MCP, in Joeven’s terms?

    • A new LLM trained by a standards body
    • A standard tool/resource transport between a host and servers
    • A guarantee that tools are safe
    • A vector database

19. MCP Host Policy

    19. If an MCP server exposes delete_repo, who must block it for a support bot?

    • Nobody — MCP is safe by spec
    • The host allowlist, before the tool is advertised or executed
    • The model, if you ask nicely
    • DNS

20. Permissions and Least Privilege

    20. Where should least privilege be enforced?

    • Only in the system prompt, as a polite request
    • In runtime policy: allowlists, argument bounds, and tight credentials behind tools
    • In the marketing site
    • By using more adjectives

21. Human Approval

    21. What should happen if a human never answers an approval?

    • Run the refund anyway
    • Treat timeout as deny; do not execute
    • Ask the model to approve itself
    • Retry 100 times

22. Path and Host Allowlists

    22. When should you check that a URL host is allowed?

    • After you download the page
    • Before any network call
    • Only if the model looks nervous
    • In the CSS

23. Identity and the Confused Deputy

    23. Whose permission should a refund check?

    • Only the cloud key on the server
    • The end user’s authority, then the tight bot credential as a second fence
    • The system prompt
    • Whoever spoke last in the chat

24. Tool Fixtures and When Tools Fail

    24. What is the cheapest way to test a tool schema?

    • Hire a larger model
    • A fixture list: valid, missing field, extra key — run in CI
    • Only production traffic
    • Reading the README once