If you’re building on AWS Bedrock, there’s a big shift happening in how agents get built. Here’s the breakdown. 𝗧𝗵𝗲 𝗢𝗻𝗲-𝗟𝗶𝗻𝗲𝗿
→ Bedrock Agents (Classic): a fully managed, opinionated agent service. AWS controls the orchestration loop. → Bedrock AgentCore: a modular infrastructure platform. You bring your own agent code and control the orchestration loop yourself.
𝗕𝗲𝗱𝗿𝗼𝗰𝗸 𝗔𝗴𝗲𝗻𝘁𝘀 (𝗖𝗹𝗮𝘀𝘀𝗶𝗰) Launched November 2023, now rebranded “Classic” and moving into maintenance mode — closed to new customers after July 30, 2026.
What you get:
AWS-managed ReAct-style orchestration loop
Action Groups (OpenAPI schemas or Lambda functions) for tool calls
Built-in RAG via Knowledge Bases
Managed session memory (short-term, within a conversation) Launched in preview May 2026, GA June 2026 — this is AWS’s forward direction for agent infrastructure.
You bring your own agent code — any framework: Strands Agents, LangGraph, CrewAI, LlamaIndex, or a fully custom build — and AgentCore supplies the surrounding infrastructure through 12 components:
Runtime — deploys your agent in serverless, session-isolated containers with auto-scaling
Harness — an optional thin orchestration layer if you want AWS to manage the loop
Memory — persistent, long-term state across sessions (not just session-scoped) Gateway — managed tool connectivity with auth and rate limiting
Identity — OAuth2/OIDC auth for tools, with token refresh and per-user permissions
Code Interpreter — sandboxed code execution
Browser — web browsing capability
Observability — real-time tracing, logging, and metrics
Payments — support for agent-initiated transactions
Evaluations — testing and scoring against datasets
Policy — guardrails and access control
Registry — a catalog for discovering and reusing agents
𝗞𝗲𝘆 𝗗𝗶𝗳𝗳𝗲𝗿𝗲𝗻𝗰𝗲𝘀 𝗮𝘁 𝗮 𝗚𝗹𝗮𝗻𝗰𝗲
Philosophy: Classic handles everything for you. AgentCore hands you the infrastructure and lets you own the agent logic.
Orchestration: Classic is fixed and AWS-managed. AgentCore is fully yours to write, in any framework.
Models: Classic is Bedrock-only. AgentCore supports any model, any provider.
Memory: Classic is session-scoped. AgentCore is persistent and cross-session.
Tools: Classic uses Action Groups. AgentCore uses a Gateway with managed OAuth and rate limiting.
Multi-agent: Limited in Classic. First-class in AgentCore. Observability: Classic relies on CloudWatch. AgentCore has a dedicated tracing and metrics component.
Evaluation: Not built into Classic. Native in AgentCore.
Learning curve: Classic — about 30 minutes to a first agent. AgentCore — steeper, since you need to understand each component.
Status: Classic is entering maintenance mode. AgentCore is under active development and now GA.
𝗪𝗵𝗲𝗻 𝘁𝗼 𝗨𝘀𝗲 𝗪𝗵𝗮𝘁
Bedrock Agents Classic fits best when:
The current setup is already working well
The flow is simple: prompt → tool calls → response
No custom orchestration logic is needed
Bedrock-only models are sufficient
Zero infrastructure management is the priority
AgentCore fits best when:
Custom orchestration or multi-step planning is required
Non-Bedrock models, or a mix of models, are needed
The system involves multiple agents
Persistent memory across sessions is required
Fine-grained, per-user tool authentication is needed
A framework like LangGraph, CrewAI, or Strands is in use
Production-grade observability and evaluations are required
The deployment is for enterprise use with VPC/PrivateLink requirements
𝗧𝗵𝗲 𝗧𝗮𝗸𝗲𝗮𝘄𝗮𝘆
Bedrock Agents Classic was AWS’s “easy button” for agents. AgentCore is the platform for teams that have outgrown that button and want real control over orchestration, memory, and infrastructure — without giving up managed deployment, observability, and security.