Here are three simple, architectural rules of thumb to help you choose the right path, along with a look at how to bridge the gap between custom local code and secure cloud production.
Before you write a single line of state-management code, ask yourself: Does my agent need custom orchestration logic, or can it be defined by instructions and schema-driven tools?
If you want a fully managed experience where the cloud provider handles the underlying infrastructure, scaling, and orchestrating loop, Amazon Bedrock Agents is your starting point. With Bedrock Agents, developer effort is shifted from writing boilerplate orchestrators to configuring action groups and integrating knowledge bases. You define the agent's instructions, hook up your APIs, attach your guardrails, and let Bedrock coordinate the reasoning loops behind the scenes. It has a low learning curve and requires zero infrastructure provisioning, making it ideal for teams that want production-ready capabilities without a massive development overhead.
When to use it:
If a managed configuration is too restrictive and you need to write custom code, your next decision is about how your agents should coordinate. For Deep AWS-Native Integration: Strands Agents
If your organization is heavily invested in AWS and you want custom code that integrates natively with the cloud ecosystem, Strands Agents is a premier choice. It offers the strongest AWS integration along with excellent support for foundation models like Anthropic Claude and Amazon Nova. It balances powerful autonomous workflow capabilities with moderate learning curves, making it an excellent default for enterprise-scale developers on AWS.
If your application requires a highly complex, custom state machine with cyclical loops, branching logic, and precise control over the conversation flow, LangGraph (and LangChain) is the gold standard. It is unmatched in workflow complexity and multimodal support, though it comes with a steep learning curve.
For Role-Based Collaboration: CrewAI
If your project depends on role-playing, where multiple specialized agents (like a researcher, a writer, and a code reviewer) must collaborate asynchronously to execute a sequential pipeline of tasks, CrewAI is a natural fit. It simplifies team-oriented architectures, allowing you to orchestrate collaborative agent "crews" with clear roles. A common mistake is assuming that choosing an open-source framework like CrewAI or LangGraph means you have to build, scale, and secure your own hosting infrastructure from scratch.
This is where Amazon Bedrock AgentCore enters the picture. Think of AgentCore as the production bridge. It is an agentic platform designed to host, secure, and operate agents built on any framework—whether it is LlamaIndex, CrewAI, LangGraph, or custom code.
AgentCore eliminates the undifferentiated heavy lifting of managing agent infrastructure. It provides several modular, serverless services that you can use together or independently:
AgentCore Runtime: A secure, isolated, and scalable serverless environment to run your custom agent containers without managing servers.
AgentCore Memory: A managed memory system that handles short-term and long-term context retention so your agents can build personalized, coherent conversations over multiple sessions.
AgentCore Gateway: A secure service that converts standard enterprise APIs into Model Context Protocol (MCP) tools, making them instantly discoverable and usable by your agents.
AgentCore Identity: Assures compliance and security by assigning unique, verifiable identities to agents, enabling fine-grained access control to enterprise data.
AgentCore Observability: Real-time tracing and monitoring of agent execution paths, giving you deep visibility to audit reasoning and debug failures in production.
By choosing a custom framework for development and pairing it with Bedrock AgentCore for production deployment, you get the absolute best of both worlds: ultimate open-source flexibility combined with enterprise-grade security, scalability, and managed operations.
Summary: Your Decision Matrix