AWS Deprecates Bedrock Agents Classic in Shift Toward Modular AgentCore Architecture

AWS has moved Amazon Bedrock Agents into maintenance mode as Amazon Bedrock Agents Classic, closing onboarding to new customers and establishing Amazon Bedrock AgentCore as its primary platform for building and operating AI agents. Under this architectural migration path, existing Bedrock Agent Classic workloads remain operational with ongoing security maintenance, while all new capabilities, multi-agent coordination features, and runtime scaling are centered entirely on AgentCore. The transition introduces a modular architecture comprising dedicated serverless runtime execution, Model Context Protocol (MCP) integrations, decoupled episodic memory services, and externalized policy evaluation. This shift directly impacts cloud architects, platform engineers, and AI developers who previously relied on Bedrock's turn-key, configuration-based agent harness. Bedrock Agents Classic packaged orchestration loops, tool calling via Lambda action groups, and retrieval-augmented generation (RAG) into a single opinionated service. While effective for rapid prototyping, this tightly coupled design created severe bottlenecks for complex, multi-step business logic requiring cross-agent collaboration, state persistence across extended sessions, or framework flexibility. By replacing an all-in-one abstraction with AgentCore's modular infrastructure primitives, AWS is acknowledging that enterprise agentic systems require the same architectural decoupling as modern microservices. The evolution reflects a broader consensus across the cloud and AI landscape: the initial wave of proprietary, closed-loop agent builders cannot accommodate diverse production demands. Enterprise engineering teams increasingly demand framework-agnostic runtime infrastructure that works interchangeably with open-source orchestration libraries—such as LangGraph, CrewAI, AutoGen, and Strands—while integrating deterministic governance layers. In this model, security policy enforcement is shifted from prompt-level instructions to low-latency gatekeepers (such as Cedar-backed deterministic authorization), and memory is treated as a shared state service rather than an ephemeral prompt artifact. In practice, practitioners must audit existing Bedrock Agent Classic configurations and evaluate migration strategies toward the AgentCore runtime. Teams should decouple their business logic from proprietary action group schemas, migrating tool interfaces toward standard protocols like MCP. Furthermore, platform engineers should take advantage of AgentCore's independent identity layer and persistent episodic memory to avoid session state loss in multi-agent handoffs. While migrating necessitates additional configuration overhead compared to the old low-code harness, the resulting decoupling gives engineering teams greater resilience, enhanced observability through structured tracing, and the freedom to swap underlying foundation models without rewriting orchestration pipelines.