Strands Agents

1 · What it is

Strands Agents is a free toolkit from AWS for making AI agents in Python or TypeScript. An AI agent is a program where a language model (the AI that reads and writes text) can also take actions, such as reading a file or asking another service for data through an API (a set of rules programs use to talk to each other). Then it looks at what happened. Strands is open source under the Apache 2.0 licence, so anyone can read and reuse the code. AWS announced it on 16 May 2025.

Strands says an agent needs three things: a model, some tools and a prompt (your written instruction). Some frameworks make you define complex workflows for the agent. With Strands you mostly write the prompt and list the tools. The model plans its own steps and picks which tools to use. AWS says newer models can already reason and use tools. That is why it chose this “model-driven” style.

Under the hood is the agent loop. Strands sends the model the prompt, the conversation so far and a short description of each tool. The model either answers in plain text or asks for tools. Strands runs them, adds the results to the conversation and calls the model again. Because the history grows each pass, the model can see what it has already done. The loop stops when the model says it is finished. It can also stop earlier if it hits a limit, such as a cap on turns.

The official quickstart shows this with a small puzzle. You write a Python function, letter_counter, and put @tool above it to make it a tool. You also add file_editor, a tool that comes with Strands. Then you ask how many letter R’s are in “strawberry” and to save the answer in answer.txt. The model gives letter_counter the job of counting and gives file_editor the job of saving. Nobody wrote that choice in code.

Tools can be your own functions, prebuilt tools, or any MCP server (a program that offers tools through the Model Context Protocol). The same agent code can run on Amazon Bedrock (the default), Anthropic, OpenAI, Google, or Ollama, which runs a model on your own computer with no API key. Built-in controls such as turn limits and token budgets (a cap on how much text the model may read and write) cap how much one request can use.

There are limits. The model can still pick the wrong tool, usually because a tool’s description is unclear. Tools run with the same permissions as your program, so check what each one can touch. Strands is a library inside your own code, not a hosted service, so you still deploy and run it yourself.