An AI agent is software that uses a foundation model to work toward a goal: it plans the steps, calls tools and systems to act, checks the results, and keeps going until the task is done. A single prompt produces one answer; an agent can carry out a whole process, such as handling a claim from first contact to payout. This lesson covers when an agent is the right design, where it pays off in business, and what AWS offers to build one today.
What an Agent Adds
A plain model call reads a prompt and writes text. An agent wraps that model in a loop with extra abilities:
- Planning: breaking a goal into smaller, ordered steps.
- Tool use: calling APIs, databases or other software to look things up and make changes, such as checking an order status or booking a slot.
- Memory: keeping context within a conversation and, where needed, across sessions.
- Autonomy: deciding the next step itself based on what each tool returned, without a person directing every move.
The value of an agent is action. If a task only needs an answer, an agent is extra cost and delay.
Business Applications
Agents fit work that spans several steps and systems:
- Customer service: look up an account, check an order, issue a replacement, or escalate to a person.
- Claims and requests: gather missing details from the customer, check policy rules, and file the claim.
- Back office: flag invoices with missing data, reconcile records, or prepare reports from several sources.
- Operations: monitor equipment data and open maintenance tickets before something fails.
The business benefits are productivity, because staff hand off repetitive multi-step work; lower costs from fewer manual errors; and faster responses for customers.
When Not to Use an Agent
An agent's loop means several model calls per task, more tokens, more latency and more things that can go wrong. If the job is answering questions from documents, a RAG knowledge base alone is simpler and cheaper. If the steps are fixed and never vary, a regular coded workflow is more predictable. Use an agent when the steps depend on what happens along the way.
Keeping Agents Safe
Because agents take real actions, design for control:
- Give each agent only the tools and permissions its job needs.
- Require human approval for high-impact actions, such as large refunds or account closures.
- Log every step so actions can be reviewed and audited.
Building Agents on AWS
Amazon Bedrock AgentCore is AWS's current platform for building, deploying and operating agents securely at scale. Its services can be used together or separately and work with any framework or foundation model. They include a serverless Runtime for hosting agents, Memory, a Gateway that turns existing APIs into tools, Identity, Policy controls on tool calls, and Observability for tracing each step.
The older Amazon Bedrock Agents feature, now called Bedrock Agents Classic, is no longer open to new customers. Existing users can keep using it, but AWS points new projects to AgentCore.
On the Exam
Choose an agent when a scenario involves multiple steps, calls to other systems, or actions taken on the user's behalf. Choose plain RAG when the task is only answering questions from documents. For a new agent project on AWS, the answer is AgentCore.