What is Strands Agents?

Last week, we took a look at the Model Context Protocol and how it creates a foundation for the building of autonomous AI agents. MCP Servers provide complete sets of tools and integrations that can enable highly contextual workflows. After spending a few days working with MCP in Python using FastMCP, I was able to create an agentic integration to Home Assistant, which is my preferred home automation platform. 

But, what if I want to go above and beyond the creation of MCP servers to create multi-agent workflows?

Introducing: Strands

Late last week, AWS announced a new open source SDK for AI Agents called Strands. Where FastMCP and similar frameworks can be used to build MCP servers, Strands gives developers the ability to create agents using the LLM of their choice, and to chain and link these agents together to solve complex problems. Strands also contains native support for MCP, allowing it to leverage third-party MCP Servers in Strands agents.

In order to more deeply understand Strands, I decided to build a multi-agent workflow using a combination of third-party MCP tools and my own MCP Server for Home Assistant. Given an agent for research, an agent for planning, and an agent for taking action, I can start by asking for a user’s input and then chain these agents together to accomplish a goal.

I’d like to support the following workflow:

1. The user will be prompted to enter the name of a city in the United States
2. The user’s city will be given to a “research agent” who is tasked with providing the following information about the specified city:
   1. Is the city’s population more than a million?
   2. Is the city a state capital?
   3. Is the city’s average temperature lower than 60F?
3. The facts will be sent to a “planning agent,” which will evaluate the research findings against specific rules and provide an ordered action plan:
   1. If the population of the city is more than a million, then turn my office ceiling fan light to 10% and turn off the office ceiling fan.
   2. If the city is a state capital, turn off my office air conditioner.
   3. If the city’s average temperature is lower than 60F, set the office ceiling fan light to 30%
   4. If the city’s average temperature is higher than 10F and the city is a state capital and the population is more than a million, close my office blinds.
4. Finally, the action plan will be handed to an “action agent”, which has access to my Home Assistant MCP Server.

My Experience Using Strands

After a few hours of tinkering, I was able to get this workflow executing reliably. While this is a contrived example, it does demonstrate that Strands makes it straightforward to implement deeply-linked multi-stage workflows that run autonomously.

Modern AI Workflows are going to be increasingly addressed by using multiple ML models and a combination of in-house developed agents and third-party MCP Servers. Rather than thinking about “one model to rule them all,” architecture and solutions design will require thinking about orchestrating solutions that take advantage of independent, purpose-built agents. Tools like Strands and Bedrock provide us with the ability to build extremely sophisticated automation flows that can leverage any model, so I would encourage you to take a look at Strands.