The names in this area sit at different layers and get read as alternatives. What makes an application agentic in the first place comes first; the choice between frameworks goes deeper than this level needs. This card is the one framework AWS publishes.
Flash card
Strands Agents: the open-source AWS framework for building agents in code.
- It is model-driven. Instead of an if-then chain you write at build time, you describe the job in a prompt and hand the model a set of tools. The framework runs the loop and the model produces the plan at run time.
- It is open source under Apache 2.0 and ships as a Python and a TypeScript library, so it runs wherever those run: a laptop, an AWS Lambda function, a container on AWS Fargate, or Amazon Bedrock AgentCore Runtime. The agent definition is the same in each; what changes is the entrypoint wrapper, a Lambda handler for instance.
- It supports the Model Context Protocol [MCP], the common way to connect an agent to external systems. A tool exposed over MCP is reachable by any MCP client, so an internal delivery-lookup server gets written once rather than per agent.
- It is the code-first option next to the managed harness in Amazon Bedrock AgentCore, where you declare the model, the tools and the instructions in configuration instead of writing the loop. The harness is itself built on Strands. The earlier service, Amazon Bedrock Agents, is now Amazon Bedrock Agents Classic: in maintenance mode from 30 July 2026, its model catalogue frozen at that date, and agent creation closed to accounts with no prior usage, so a new team cannot pick it.
- Being a library rather than a service, it brings no managed infrastructure. It persists session state to a file or an S3 bucket on its own; hosting, memory that survives across sessions and identity for an agent acting on somebody’s behalf come from Amazon Bedrock AgentCore, along with session isolation and a gateway that turns existing APIs into MCP tools.
Pick it when
Pick it when the team wants agent behaviour written, versioned and tested like application code, when the tools are internal services somebody has to wire up anyway, and when the agent may move hosts later.
It's the wrong answer when
It is the wrong fit for a team that would rather declare an agent in configuration than own the loop, where the AgentCore harness serves, and for a business analyst who wants answers over the organisation data, where Amazon Quick serves.