Agentic AI: AIF-C01 study notes and practice questions
What to know about Agentic AI for the AIF-C01 exam, from the Fundamentals of Generative AI domain (24% of the exam guide), then 4 free questions to test it. Part of the AIF-C01 study guide.
Agentic AI was added in version 1.1 of the exam guide. Questions ask what makes something an agent, how agents use tools and memory, and which AWS service fits.
Chatbot vs agent
A generative AI chatbot answers questions. An AI agent works toward a goal: it plans steps, decides which tools to call (APIs, databases, other systems), acts, checks the result and continues. “Rebook a passenger’s canceled flight” needs an agent; “explain the rebooking policy” needs only a chatbot or RAG.
Building blocks
Piece
Job
Foundation model
Reasons and plans
Tools / actions
APIs and functions the agent can call
Memory
Short-term memory keeps the current conversation; long-term memory keeps facts such as preferences across sessions
Knowledge
Documents the agent can retrieve (RAG)
Orchestration
The loop of plan, act and observe
Guardrails and policy
Limits on what the agent may say and do
Model Context Protocol (MCP)
MCP is an open protocol for connecting AI applications to tools and data sources in a standard way, so an integration built once works with any MCP-compatible agent.
Multi-agent patterns
A supervisor agent can route work to specialist collaborator agents, each with its own tools and permissions. This keeps each agent focused and limits what each can access.
AWS services
Amazon Bedrock Agents: managed agents that call your APIs and knowledge bases.
Amazon Bedrock AgentCore: services to run agents in production with any framework or model: Runtime (hosting), Memory, Gateway (turn APIs into tools), Identity, Observability, and Policy (rules enforced outside the agent on every tool call).
Strands Agents: an open-source SDK, first released by AWS, for building agents in code.
Common traps
Calling any chatbot an agent. Agents take actions with tools.
Relying on the prompt to stop risky actions. Enforce rules outside the model, with policies and least-privilege permissions.
A travel company already runs a generative AI chatbot that answers questions about its booking policies.
It now wants a system that, given the goal 'rebook this customer's canceled flight', decides its own steps, calls the airline and payment APIs, checks each result and retries failed calls.
Which kind of AI describes the new system?
Choose one answer.
AA. Agentic AI that plans steps and uses tools to act towards a goal.
BB. A computer vision model that classifies photos of boarding passes.
CC. A clustering model that groups customers by their travel history.
DD. A larger generative AI chatbot that writes longer and more detailed answers.
Show the answer and explanation
Answer: A
A. Right: Correct. An agent breaks a goal into steps, calls external tools and APIs, and checks the results. That matches the rebooking system the company wants. Learn more: What are AI agents? (opens in a new tab)
B. Wrong: Computer vision classifies images. It could read a boarding pass, but it does not plan a rebooking or call booking systems.
C. Wrong: Clustering groups similar records without labels and does not act on them. It cannot rebook a flight or take a payment.
D. Wrong: A chatbot that writes longer answers still only generates text. It cannot call APIs or take actions toward the customer's goal.
Explanation
A generative AI chatbot produces answers, while an AI agent plans steps and uses tools to reach a goal. The rebooking system must decide what to do, act through APIs and check outcomes, so it is agentic. This question tests recognizing agentic behavior from the description of what the system must do.
A software company has built five AI agents for support, code review and release planning.
Each agent needs the same ticketing system, code repository and internal wiki, and today every agent has its own custom connector to every tool. The team wants one open protocol so each tool is connected once and any compatible agent can use it.
Which standard addresses this?
Choose one answer.
AA. Model Context Protocol (MCP), an open protocol for connecting AI applications to tools and data sources.
BB. Retrieval Augmented Generation (RAG), which adds retrieved documents to the prompt at query time.
CC. Reinforcement learning from human feedback (RLHF), which tunes a model using ratings from human reviewers.
DD. Open Neural Network Exchange (ONNX), which moves trained models between frameworks and runtimes.
B. Wrong: RAG supplies relevant documents to a model at query time. It does not define how agents connect to ticketing systems or repositories.
C. Wrong: RLHF is a training method that aligns a model with human preferences. It does not connect agents to tools.
D. Wrong: ONNX is a format for moving trained models between frameworks. It does not standardize how agents call external tools.
Explanation
The Model Context Protocol is an open protocol that connects LLM applications to external data sources and tools, so integrations can be built once and reused. RAG, RLHF and ONNX each solve a different problem in the AI stack. This question tests recognizing the standard for agent-to-tool connections.
A shoe retailer's shopping assistant should remember that a returning customer prefers a certain brand and size, even when the customer comes back weeks later in a new chat.
The team wants only the key facts kept, not every message.
Which agent capability provides this?
Choose one answer.
AA. A higher temperature so the agent gives more varied, creative replies.
BB. A much larger context window, so the current conversation can be longer.
CC. Loading the customer's full past chat transcripts into every new prompt, with no summary.
DD. Long-term memory that stores key facts, such as preferences, across sessions.
Show the answer and explanation
Answer: D
A. Wrong: Temperature changes how random each reply is. It does not store anything about a customer between chats.
B. Wrong: A larger context window holds more text within one conversation. It is lost when the session ends, so it cannot recall preferences weeks later.
C. Wrong: Full transcripts can carry facts across sessions, but they fill the prompt and cost more on every call. The team wants only the key facts kept, not every message.
D. Right: Correct. Long-term memory extracts important facts and preferences from conversations and keeps them across sessions. The assistant can then recall the brand and size in a later chat. Learn more: Amazon Bedrock AgentCore Memory (opens in a new tab)
Explanation
Keeping facts across sessions requires memory that lives outside a single conversation. Long-term memory stores key facts such as preferences, while short-term memory covers only the current session. Temperature and context window size do not store anything between chats. This question tests distinguishing memory across sessions from other settings.
A bank wants one assistant for mortgage questions, card disputes and general account queries.
Each area uses different tools, and only its own team may access that area's data. The architects want a specialist agent for each area, with one component that decides which specialist takes each request and combines the answers into one reply.
Which design pattern is this?
Choose one answer.
AA. A batch inference job that processes all customer requests overnight and writes results to storage.
BB. A single agent with one very long system prompt covering every area, tool and set of data permissions.
CC. A supervisor agent that routes requests to specialist collaborator agents and combines their results.
DD. Separate chatbots for each area, with the customer choosing which bot to open and no shared answer.
Show the answer and explanation
Answer: C
A. Wrong: Batch inference runs offline over stored data and gives no live reply. It cannot route a customer's question to a specialist in real time.
B. Wrong: One agent with every tool and every rule mixes the areas together and breaks the per-team data limits. The prompt grows hard to manage as areas are added.
D. Wrong: Separate chatbots meet the specialist and access rules, but they lose the single assistant the bank wants. The customer would have to pick the right bot, and no component combines the answers.
Explanation
A supervisor and collaborator pattern splits one assistant into specialists, each with its own tools and data permissions, and a supervisor that routes requests and combines results. Bedrock's multi-agent collaboration implements this, with the supervisor planning and routing work to collaborators. This question tests recognizing the pattern from the stated roles and constraints.
Want more on Agentic AI? The 5 full-length practice sets have 19 more Agentic AI questions, 325 in all, none of them in the free exam. One payment of $14.99, no subscription. See the practice sets →