Search⌘ K
AI Features
Home
Courses
Introduction to Generative AI and Amazon Bedrock

What Is an AI Agent?

Explore the concept of AI agents and how they extend foundation models into autonomous systems using Amazon Bedrock. Understand the ReAct reasoning pattern, core architectural components, defensive design patterns, and how to calibrate agent autonomy for complex tasks. By the end, you will grasp how to design, debug, and deploy multi-step AI workflows with managed orchestration in cloud environments.

We'll cover the following...

A single-turn LLM call is stateless and one-shot. You send a prompt, receive a response, and the interaction ends. The model cannot browse the web, query a database, retry a failed step, or iterate on its own output. For many use cases, this is sufficient. When a task requires multiple steps, branching decisions based on intermediate results, and calls to external tools, APIs, or data sources, a single model call is usually not enough.

An AI agent overcomes this limitation by wrapping a foundation model in a continuous loop. The agent observes its environment, thinks about what to do next, acts by invoking a tool, and then observes the result of that action. This cycle repeats until the agent achieves its goal or reaches a termination condition. The result is a system capable of multi-step task completion, dynamic tool use, and goal-directed behavior that no single inference call can replicate.

Amazon Bedrock provides the managed infrastructure for building foundation-model-powered applications on AWS, and Amazon Bedrock AgentsA fully managed capability within Amazon Bedrock that handles the orchestration loop, tool routing, and state management for AI agents without requiring developers to build custom plumbing. enable engineers to deploy agentic systems without constructing the orchestration logic from scratch. This lesson establishes the mental models you need to design such systems. By the end, you will understand how agents differ architecturally from basic inference, the ReAct reasoning pattern, the four core components of any agent system, agent-specific failure modes with defensive patterns, and how to choose the appropriate level of autonomy for a given task.

The following diagram visually contrasts these two paradigms:

Single-turn LLM inference vs. AI agent loop
Single-turn LLM inference vs. AI agent loop

With this architectural distinction clear, the next step is to understand the reasoning framework that powers the agent’s decision-making in each iteration.

The ReAct pattern explained

The dominant framework for LLM-based agents is ReAct (Reasoning + Acting)A prompting and orchestration pattern where the model alternates between generating internal reasoning traces (Thought), executing external actions (Action), and processing returned results (Observation) in a repeating cycle.. This pattern structures each iteration of the agent loop into three discrete phases:

  • Thought: The model reasons internally about what to do next, producing an explicit chain of logic such as “I need to list all S3 buckets and check their encryption configuration.”

  • Action: Based on that reasoning, the model invokes a specific tool, for example, calling the ListBuckets API followed by GetBucketEncryption.

  • Observation: The model processes the tool’s returned result and decides whether the goal is met or another iteration is required.

This structure makes agent execution easier to inspect and ...