Strands Agents: building AI agents with the AWS SDK

An SDK born inside Amazon

AWS introduced Strands Agents in May 2025 as a tool that came out of its own teams, who were using it internally to build AI agents. Its philosophy is “model-driven”: instead of describing every step in a graph, you give the agent instructions, a model and a list of tools, and the model decides when to call which tool (tool calling). The SDK handles the rest: the execution loop, turn and token limits, streaming, structured output, memory, sessions and tracing.

In 2026 the project was reorganized into a single repository that holds the Python SDK, the TypeScript SDK, a command-line tool and the “Strands harness”, a ready-to-use, pre-assembled agent.

Who is it for, and what for?

Strands is aimed at developers who would otherwise write their own agent loop. Typical uses: an internal assistant that queries company APIs, an analysis agent that chains searches and calculations, or a multi-agent system where several specialists split a task. The AWS connection is a plus for teams already on Bedrock, but the SDK stays provider-neutral and connects to MCP servers.

Getting started

You need Python 3.10 or later, or Node.js 22 or later. The fastest route is to install the harness package and call a single agent-creation function. If you want full control, install the strands-agents SDK instead, create an agent object with your tools, then configure the model provider and its credentials (an API key or AWS access). The official quickstart walks through each provider.

Strengths and limitations

Strengths: a short API, two languages, model independence, production features built in (tracing, guardrails, evaluations) and a permissive license.

Limitations: the project moves fast and its structure has changed (renamed repositories, a new harness), so some tutorials may be out of date. Many examples assume an AWS account. As with any agent, every run consumes tokens: estimate the bill with the agent cost calculator.

Alternatives

LangGraph is a better fit if you want to define each step precisely, the OpenAI Agents SDK and Google ADK fill the same role for other model makers, and Pydantic AI focuses on strict typing in Python.