The AIP-C01 exam guide names Strands Agents, AWS Agent Squad, and MCP (Model Context Protocol) in its agent skills (2.1.1, 2.1.6, 2.1.7, 2.5.5) and its retrieval skill 1.5.6. Bedrock Agents covers the fully managed path; this post covers the open-source, code-first tools and the standard protocol agents use to reach their tools, and which one an exam scenario is pointing to.
Managed vs Code-First: Three Options
| Factor | Bedrock Agents | Strands Agents | AWS Agent Squad | |--------|----------------|----------------|-----------------| | Form | Fully managed service | Open-source SDK (Python, TypeScript) | Open-source framework (Python, TypeScript) | | Core idea | Configure action groups and knowledge bases | Define model, tools, and prompt; the model plans | A classifier routes each request to the best agent | | Runs on | AWS-managed runtime | Lambda, Fargate, EKS, AgentCore Runtime, and more | Inside your application | | Exam phrasing | "Managed agent with minimal code" | "Fine-grained control over agent behavior in code" | "Route requests among specialized agents" |
Strands Agents
Strands Agents takes a model-driven approach: you define a model, a list of tools, and a system prompt, and the model decides which tools to call and in what order inside the agent loop. Tools are Python functions with the decorator; the name, type hints, and docstring become the tool specification.
Validate input inside the tool and return a clear reason when it fails, so the model can correct itself. For multiple agents, Strands supports agents as tools, swarm (agents hand work to each other), graph (defined nodes and conditional edges), and workflow (a predefined sequence). Session managers persist conversation state to local files or S3, and every decision in the agent loop is traced and can be exported with OpenTelemetry.
AWS Agent Squad
AWS Agent Squad, formerly Multi-Agent Orchestrator, sits in front of specialized agents. A classifier reads the user input, the agent descriptions, and the conversation history, then picks the best agent; the orchestrator stores the exchange so context carries over when the topic changes. A SupervisorAgent can call team agents in parallel and merge their answers. The managed alternative is Bedrock Agents multi-agent collaboration.
MCP: A Standard Way to Reach Tools
MCP is an open protocol that standardizes how AI applications connect to tools and data. A host (an agent or IDE) creates one MCP client per server; servers expose tools, resources, and prompts over JSON-RPC 2.0, using stdio locally or Streamable HTTP remotely. Build a tool server once and every agent can reuse it.
Skill 2.1.7 asks where to run MCP servers on AWS: Lambda for stateless, lightweight tools; Amazon ECS on Fargate for complex, long-running tools or heavy dependencies; and Amazon Bedrock AgentCore Gateway to expose existing APIs and Lambda functions as MCP tools without rewriting them. Exposing Knowledge Base retrieval as an MCP tool gives every agent the same way to search company knowledge (1.5.6).
!One agent, many tools through MCP
Running Tools Reliably
Validate parameters and return clear error messages. Use Lambda timeouts and idempotent operations. Give each tool its own least-privilege IAM role. Require approval for risky actions: Return of Control in Bedrock Agents, or a Step Functions callback for whole workflows. Track call counts, failure rates, latency, and delegation between agents against a baseline.
Where AgentCore Fits
Amazon Bedrock AgentCore provides managed building blocks for running agents built with any framework: Runtime (serverless execution with session isolation), Memory (short-term and long-term), Gateway (APIs and Lambda as MCP tools), Identity (credentials for outbound access), and Observability (OpenTelemetry traces in CloudWatch). The skill statements don't name it, but it appears naturally when you take a Strands and MCP agent to production.
Exam Quick Reference
"Managed agent with minimal code" -- Amazon Bedrock Agents "Code-first agent: define model, tools, and prompt" -- Strands Agents "Turn a Python function into a standard tool" -- Strands @tool decorator "Route each request to the best specialized agent" -- AWS Agent Squad classifier "Lead agent calls team agents in parallel" -- Agent Squad SupervisorAgent "Standard tool interface reused by many agents" -- MCP "Stateless, lightweight MCP tools" -- Lambda MCP server "Complex, long-running MCP tools" -- Amazon ECS MCP server "Expose existing APIs as MCP tools without code changes" -- AgentCore Gateway "Risky action needs human approval" -- Return of Control or a Step Functions approval step