AI agents — AIF-C01

What it is

An AI agent is a software program that interacts with its environment, collects data, and uses that data to perform self-directed tasks in pursuit of a predetermined goal. Humans define the objective; the agent autonomously determines the sequence of actions needed to achieve it.

On AWS, Amazon Bedrock Agents lets you build and configure autonomous agents that orchestrate interactions between foundation models, data sources, software applications, and user conversations — automatically calling APIs and invoking knowledge bases to complete tasks.

Mental model

Think of the difference between a vending machine and a personal assistant.

A vending machine (a standard model invocation) takes your input, runs a fixed process, and returns an output. Done.

A personal assistant (an AI agent) receives your goal, figures out what steps are needed, gathers missing information by asking follow-up questions, calls the right tools in the right order, checks whether the outcome matches your goal, and adjusts if it does not. The loop keeps running until the goal is met or the agent determines it cannot proceed.

The defining characteristic is the agent loop: perceive → reason → act → observe → repeat.

When to use it

Scenario	Right tool	Why
User asks a one-shot question; answer is in the model's training data	Direct foundation model invocation	No tools or multi-step reasoning needed
Answer requires retrieving current private documents	RAG / Knowledge Base	Retrieval augments a single generation step; no action execution needed
Task requires calling external APIs, executing multi-step workflows, or decisions across multiple interactions	AI agent	Agent orchestrates tool calls, handles branching logic, tracks state across steps
Multiple specialized sub-tasks must coordinate	Multi-agent system	Orchestrator agent delegates to specialist agents

Common misconception

Agents are not just a smarter chatbot that calls one tool.

The exam-exploitable trap is treating an agent as a single round-trip: user sends a message → model picks a tool → tool runs → answer returned. That describes a tool-augmented completion, not an agent.

What distinguishes an agent is the orchestration loop. According to AWS documentation, agents extend foundation models to understand user requests, break tasks into smaller steps, collect additional information from users through natural conversation when needed, and make sequences of API calls — evaluating each step's output before deciding what to do next. The reasoning process is multi-step and adaptive, not a single function call.

A related trap: assuming that because an agent uses a knowledge base it is RAG. An agent may query a knowledge base as one of many possible actions, but RAG alone (retrieve → generate) is not an agent — it lacks the planning and multi-step action loop.

How it shows up on the exam

The cognitive target is distinguishing an agent architecture from adjacent patterns (plain model invocation, RAG, and single-tool augmentation) based on what the use-case actually requires.

Signal phrases to watch for:

• "autonomously complete a multi-step task" → suggests agent
• "make API calls to company systems" → suggests agent's action groups
• "break down a complex request into steps" → suggests agent orchestration
• "retrieve documents and generate an answer" → suggests RAG, not necessarily an agent
• "without provisioning infrastructure" — Amazon Bedrock Agents manages prompt engineering, memory, monitoring, and API invocation on your behalf

A question may describe a business workflow — AWS documentation cites processing insurance claims and making travel reservations as agent use cases — and ask which architecture is appropriate. Candidates often reach for RAG when the scenario actually requires action execution and state management across multiple turns. That gap is where agent questions test understanding.

Related concepts

• Bedrock Knowledge Bases — the data-retrieval component an agent can query as one of its actions
• RAG Design Considerations — understand where retrieval-augmented generation ends and agent orchestration begins
• Retrieval-Augmented Generation — the pattern agents may incorporate but that does not by itself constitute an agent