Chapter 3: AI Agents19%（当サイト設定） · 6 topics

Q1. Which best describes the difference between AI agents and conventional conversational generative AI?

1. An agent plans toward a goal, acts using tools, observes the results and decides its next action
2. An agent accepts no human instructions at all and always sets its own goals
3. An agent does not use LLMs and runs only on rules, so it cannot execute tools or connect externally
4. An agent cannot generate text and only classifies images

Q2. Which best describes the background to the attention AI agents are receiving?

1. Because using agents makes model usage free of charge
2. Because law restricted business use of generative AI and required agents instead
3. Because research has proven that agents make no errors even without human checks
4. Because LLM reasoning abilities and tool integration mechanisms have improved, making it possible to delegate multi-step work

Q1. A chatbot about internal rules gives outdated content or nonexistent clauses. Which is the best reason to introduce RAG?

1. To search current internal documents as evidence and reduce hallucinations
2. To encrypt input and prevent leaks of internal documents
3. To redo pretraining on internal rules to update knowledge
4. To lower the temperature and stabilize wording

Q2. Which best describes chunking in RAG?

1. Splitting the user's question into words and having each word answered in order; the finer, the more accurate
2. Displaying the answer text in pieces that fit on the screen; the shorter the pieces, the more accurate the answer
3. Splitting documents into units of a suitable size for search; too large or too small lowers search accuracy
4. Splitting documents by department and storing them on separate servers; more storage locations raise search accuracy

Q1. Which best describes an AI agent's “short-term memory” and “long-term memory”?

1. Short-term memory is the context of the current conversation or task; long-term memory stores past interactions, user information and so on for later use
2. Short-term memory refers only to GPU memory and long-term memory only to hard disks
3. They are divided by who remembers: short-term is what users remember, long-term is what administrators remember
4. By rule, short-term memory is deleted automatically after a day and long-term memory after a year

Q2. An internal Q&A agent loses accuracy as conversations get longer. From a context engineering view, which fix is best?

1. Switch to a larger context window and keep feeding the whole history
2. Refine the prompt wording and give instructions more politely
3. Further train the model on internal documents to store knowledge in weights
4. Select the instructions, documents, history and tool results the task needs and put them in context

Q1. Which best describes multi-agent?

1. Adding the weights of multiple models together into one model to improve performance
2. A configuration where multiple agents with different roles divide and coordinate work on a task
3. Multiple users using one agent at the same time
4. Showing the agent's screen on multiple displays

Q2. Which statement about MCP is NOT appropriate?

1. It is positioned as a common standard connecting AI apps with external tools and data sources
2. Once connected to an MCP server, it can be used safely without setting access permissions
3. Managing which operations are allowed for each connection target becomes important
4. Once you build an MCP server, multiple compatible AI apps can use it

Q1. To improve a coding agent, you set up automatic test runs, available tools and permission scope. What is this work called?

1. Human-in-the-loop
2. Prompt engineering
3. Harness engineering
4. Fine-tuning

Q2. Which best describes deep research?

1. A function where an agent searches and reads across multiple sources and compiles a research report with citations
2. A function that digs deep into old data in internal databases and deletes it
3. A function that estimates image depth to make 3D images or models from flat photos
4. A function that spends time repeatedly reading and summarizing only one specified web page

Q1. After a file conversion error, an agent finished the task another way. Which mechanism supports this?

1. Retraining the model on the spot from user ratings
2. A loop of reasoning, acting, observing results and replanning
3. Repeating the same step a fixed number of times on errors
4. A workflow with every step fixed in advance and run in order

Q2. Which is the most appropriate way to roll out agents in work?

1. Roll out to every task company-wide at once and stop wherever problems appear
2. No need to revisit staff roles; roll out without changing workflows to reduce frontline burden
3. Do no evaluation or monitoring until results appear
4. Pilot small with low-impact routine work, evaluate effects and risks, then expand scope