What Is an AI Agent? A Simple Guide for AWS Beginners
New to AI agents? Learn what an agent is, how it works, and how Strands Agents and Amazon Bedrock AgentCore help you build and run one on AWS.
An AI agent is a program that uses an AI model to decide what to do, calls tools to do it, and keeps going until the job is done. On AWS, Strands Agents is the open source SDK you use to write the agent, and Amazon Bedrock AgentCore is the service that runs it safely in production.
The 30-second answer
- Want to build an agent? Use Strands Agents. It is a code library for Python and TypeScript.
- Want to run it for real users, with memory, security and monitoring? Use Amazon Bedrock AgentCore.
- Want a quick test? Strands on your laptop is enough. You do not need AgentCore on day one.
- Not sure which model? Strands works with Amazon Bedrock, Anthropic, OpenAI, Google, Ollama and others.
You write the agent with one, and you run it on the other. They are partners, not rivals.
What is an AI agent, in plain words?
A normal chatbot answers a question and stops. An agent can act. You give it a goal, such as "find the refund policy and answer this customer", and it works in a loop:
- Think: the AI model reads the goal and decides the next step.
- Act: it calls a tool, such as a search, a database lookup or an API.
- Look: it reads the result.
- Repeat until it has an answer.
So an agent is three parts: a model (the brain), tools (the hands) and a loop (the habit of trying again). AWS describes agents as autonomous apps that reason, use tools and keep context.
Strands Agents: how you build one
Strands Agents is an open source SDK for Python and TypeScript. The idea is that you only supply the model and the tools, and Strands runs the loop for you. A normal function becomes a tool with one decorator, and Strands reads its name, type hints and docstring to tell the model how to use it.
Here is the quickstart shape from the Strands site:
Python
from strands import Agent, tool@tooldef search_docs(query: str) -> str:"""Search the documentation."""return index.search(query) # your own search goes hereagent = Agent(model="global.anthropic.claude-sonnet-5",tools=[search_docs],)agent("Find the pricing guide")
That is a complete agent: a model, one tool, and a question. Strands also has multi-agent patterns, such as agents used as tools, Swarm, Graph, Workflow and Agent2Agent (A2A), for when one agent is not enough. It also supports MCP servers as tools.
Pick Strands when you want to write the agent logic yourself, in code you own, and run it anywhere.
Amazon Bedrock AgentCore: how you run it in production
Running an agent on your laptop is easy. Running it for thousands of users is harder: you need isolation between users, memory, safe access to tools, and a way to see what went wrong. AgentCore is a set of services for exactly that. It works with open source frameworks such as Strands, LangGraph, CrewAI and LlamaIndex, and with many models, in or outside Bedrock.
AgentCore service | What it does, in one line |
|---|---|
Runtime | A serverless place to deploy and scale your agent, with each session isolated |
Memory | Short-term memory for one chat and long-term memory across sessions |
Gateway | Turns APIs and Lambda functions into tools the agent can call (MCP) |
Identity | Manages who the agent is and what it may access |
Observability | Traces and monitoring so you can debug in production |
Code Interpreter | A safe sandbox where the agent can run code |
Browser | A cloud browser the agent can use to visit websites |
Policy | Rules that check every tool call before it runs |
Evaluations | Automatic checks of how well the agent does its job |
You do not have to use all of them. Pick the pieces you need.
Pick AgentCore when your agent has real users, needs memory or secure tool access, or must be monitored like any other production service.
Watch out: AgentCore services are billed separately, for example Runtime per second of CPU and memory used, and Memory by data stored and retrieved. Check the pricing page before you launch.
Strands vs AgentCore side by side
| Strands Agents | Bedrock AgentCore |
|---|---|---|
What it is | An open source SDK (code library) | A managed AWS platform |
Job | Build the agent | Run and protect the agent |
Where it runs | Anywhere: laptop, Lambda, containers | On AWS, in a managed runtime |
You write | Python or TypeScript | Mostly configuration and deployment |
Cost | Free library, you pay for the model | Pay for each service you use |
Which one for which situation? Real scenarios
Your situation | Choose | Why |
|---|---|---|
Learning how agents work | Strands on your laptop | Small and free to try |
A support bot for your team's docs | Strands, then AgentCore Runtime | Build first, host when it works |
An agent that remembers each customer | AgentCore Memory | Built-in short and long-term memory |
An agent that must not call risky tools | AgentCore Policy and Identity | Checks every tool call |
An agent that runs user-submitted code | AgentCore Code Interpreter | Isolated sandbox |
Several agents working as a team | Strands multi-agent patterns | Swarm, Graph and Workflow |
If two answers fit, start with Strands. You can add AgentCore later.
Common mistakes
- Starting with the platform instead of the agent. Get a small agent working in Strands first.
- Giving the agent too many tools. The model gets confused and picks the wrong one. Start with two or three.
- Skipping limits and permissions. An agent can call tools by itself, so give it only the access it needs.
- Not testing the agent. Agents can answer differently each time. Test with real examples, not just one.
- Forgetting to clean up. Delete test deployments when you finish.
Try it yourself: a 10-minute first agent
- Install Strands: run
pip install strands-agentsin a new Python environment.
- Make sure you have access to a model, for example in Amazon Bedrock, and that your AWS credentials work.
- Paste the example above, and replace the search tool with a simple one, such as a function that returns the current time.
- Run it and ask a question that needs the tool. Watch how the agent decides to call it.
- Clean up: if you tried a hosted deployment, delete it when you finish.
Sources
Facts checked against the pages above on Oct 28, 2026. Services, features and prices change often, so check the current AWS documentation before you build.
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article