QUIZ
Agents and MCP
12 questions on Agents and MCP, from 24 lessons. Pick an answer in your head, then open it.
On the roadmap for: AI Engineering.
Question 1 of 12
An agent succeeds at each step 95% of the time. Roughly how often does it complete a 20-step task with no mistakes?
- A. About 95%, the same as each step
- B. About 75%, since most steps succeed
- C. About 5%, since errors multiply fast
- D. About 36%
View answer
✓ D. About 36%
0.95²⁰ ≈ 0.36. Small per-step error rates compound over long tasks, which is why workflows or checks help.
Lesson: What Is an Agent? The ReAct Loop
Question 2 of 12
What is the core loop of a ReAct-style agent?
- A. Think, call a tool, observe, and repeat until done
- B. Ask the user for permission at every single step
- C. Answer once from memory, then stop
- D. Retrieve documents, then fine-tune on them
View answer
✓ A. Think, call a tool, observe, and repeat until done
The model reasons, acts through tools, and uses what it observes to decide the next step.
Lesson: What Is an Agent? The ReAct Loop
Question 3 of 12
Why does giving an agent 100 tools often make it worse?
- A. Each tool call is slow, so more tools mean more waiting
- B. Models can only ever call one tool per conversation
- C. Tools have to be called in the order they are listed
- D. Long tool lists eat context and make choosing harder
View answer
✓ D. Long tool lists eat context and make choosing harder
Fewer, well-described tools, or loading tools on demand, keeps selection accurate.
Lesson: Tool and Function Calling
Question 4 of 12
When does asking an agent to 'check its work' help most?
- A. Only when the answer is a single short sentence
- B. Never: models can't judge their own work at all
- C. Always, by about the same amount for every kind of task you give it
- D. When there's an outside signal, like tests or a calculator
View answer
✓ D. When there's an outside signal, like tests or a calculator
Self-review with no new information often just repeats the same mistake; verification against real feedback works.
Lesson: Planning and Reflection
Question 5 of 12
How can an agent reliably remember a user's peanut allergy 50 turns later?
- A. Hope the fact stays somewhere in the context window
- B. Save key facts to memory and bring them back when relevant
- C. Repeat the allergy in every reply to the user
- D. Switch to a bigger model with a longer context window that holds everything
View answer
✓ B. Save key facts to memory and bring them back when relevant
Context is limited and summaries lose details, so important facts are stored and retrieved deliberately.
Lesson: Memory and Context Engineering
Question 6 of 12
What is the main idea behind agent 'skills'?
- A. Procedure files loaded only when needed, saving context
- B. Giving the agent admin rights to run any program
- C. Fine-tuning a separate model for each task the agent might have to do
- D. Hard-coding every tool call in advance
View answer
✓ A. Procedure files loaded only when needed, saving context
A short description is always visible, and the full instructions load on demand, so the agent can have many skills cheaply.
Question 7 of 12
How does MCP differ from plain function calling?
- A. MCP replaces the LLM with a fixed set of tools
- B. MCP only works with one vendor's family of models, so it isn't portable
- C. They are two different names for exactly the same thing
- D. Function calling asks for a tool; MCP is how tools are offered
View answer
✓ D. Function calling asks for a tool; MCP is how tools are offered
With MCP, a tool is built once as a server and works across many agents and apps.
Lesson: MCP: The Model Context Protocol
Question 8 of 12
Your agent needs tools from your own systems and help from another company's agent. Which protocols fit?
- A. MCP for both the tools and the other company's agent
- B. Plain HTTP calls with no shared protocol at all
- C. No protocol: the two agents can share memory directly
- D. MCP for tools, and an agent-to-agent protocol like A2A
View answer
✓ D. MCP for tools, and an agent-to-agent protocol like A2A
Tools and peer agents are different connections, and each has its own protocol.
Question 9 of 12
When do multiple agents usually beat a single agent?
- A. When the work splits into parts needing separate contexts
- B. Always: more agents always means a smarter system
- C. For short, simple tasks where speed matters most
- D. Whenever you want to lower the total token cost of a long-running task
View answer
✓ A. When the work splits into parts needing separate contexts
Multi-agent setups cost many more tokens and add coordination problems, so they pay off on wide, parallel work.
Lesson: Multi-Agent Systems
Question 10 of 12
A long agent run crashes halfway after charging a customer. How do you avoid charging twice on restart?
- A. Retry the whole run from the very beginning
- B. Turn off payments for long-running agents
- C. Durable execution with checkpoints and idempotency keys
- D. Ask the model whether it already charged the customer
View answer
✓ C. Durable execution with checkpoints and idempotency keys
Recorded steps are skipped on resume, and an idempotency key makes a repeated charge a no-op.
Lesson: Agent Infrastructure: Durable Execution, Retries and Budgets
Question 11 of 12
What is the 'lethal trifecta' for an AI agent?
- A. Being slow, expensive and wrong at the same time on the same request
- B. Letting three agents talk to each other
- C. Giving the agent three tools at once
- D. Private data, untrusted content and a way to send data out
View answer
✓ D. Private data, untrusted content and a way to send data out
With all three, a hidden instruction in an email or web page can make the agent leak data. Remove at least one leg.
Lesson: Agent Safety: Permissions, Sandboxing and the Lethal Trifecta
Question 12 of 12
Why do human approvals often fail to catch an agent's mistakes?
- A. Humans are slower than agents, so they fall behind
- B. Too many requests, too little context: people rubber-stamp
- C. Approvals are usually done automatically in practice, so nobody checks them
- D. Agents hide their real requests from the reviewers
View answer
✓ B. Too many requests, too little context: people rubber-stamp
Ask for approval only on risky actions, and show what will change and why, so each review is meaningful.