Exam-style
A support team runs a single Strands agent on the Amazon Bedrock AgentCore runtime. It has split that agent into four specialists, for billing, delivery, returns and account changes, and needs a front door that routes each message to the right one and keeps the thread coherent when a subscriber changes topic halfway through. What should the team add?
Reveal the answer
C. Add Agent Squad in front as the classifier-based router, leaving the four specialists built on Strands Agents and hosted on the AgentCore runtime
The three names sit at three different layers. Strands Agents is the open-source SDK you build an agent in: model, system prompt, tools, loop. Agent Squad is an orchestration layer that goes in front of several agents. Its classifier reads the registered agents’ descriptions and the conversation so far, then routes the turn to the closest match. It carries shared conversation context, so a mid-thread topic change reaches a different specialist with the history intact. Agent Squad began in AWS Labs as Multi-Agent Orchestrator and is now maintained outside AWS by 2FastLabs, a community framework rather than an AWS service. AgentCore is the platform underneath. Its Runtime hosts what you built, isolating each session in a dedicated microVM; Memory and Identity are separate AgentCore services. Merging the four back into one agent discards the separation that made routing tractable and crowds every tool into one context window. Rebuilding the specialists on Agent Squad replaces working code for nothing. A Step Functions Choice state branches on comparison operators you write. Keyword matching is not intent classification, so a subscriber who opens with a delivery complaint and ends up closing an account lands on the wrong branch. AgentCore Memory stores conversation state; it does not select which agent runs.
Q. Four specialist agents and a front door that has to pick one per message. Strands, Agent Squad, or AgentCore?
A. Three different layers. Add Agent Squad as the classifier-based router in front; keep the four specialists built on Strands Agents and running on the AgentCore runtime.
Why? Strands builds an agent, Agent Squad routes between agents, AgentCore runs them, and MCP is one of the ways any of them reaches a tool. Adding one does not replace the others, so a routing problem gets solved by adding the routing layer rather than rewriting what already works. Note that Agent Squad has left AWS Labs for outside maintainers, which changes who supports it, not where it sits.