Amazon quietly retired its original Bedrock Agents product on July 30, 2026. The service, launched in November 2023, is now called Amazon Bedrock Agents Classic and is closed to new customers. Existing workloads continue to run, but no new features will be added and the model catalog is frozen at the maintenance mode date.

In its place, AWS is pushing teams toward Amazon Bedrock AgentCore — a fundamentally different approach to building and running AI agents at enterprise scale.

This is not a minor rebrand. For any team currently building on AWS and evaluating AI agent infrastructure, understanding the difference matters.

What Was Bedrock Agents Classic?

Bedrock Agents, launched at AWS re:Invent 2023, gave developers a managed, configuration-driven way to build agents. You defined Action Groups (via OpenAPI schemas or Lambda function definitions), attached a Knowledge Base, wrote instructions, and Bedrock handled the reasoning loop, tool invocation, and response synthesis.

It was opinionated, relatively low-code, and tightly bound to Bedrock-hosted models. For teams that wanted to ship something quickly without building custom orchestration, it was a solid starting point.

The problem is that “quickly” became “not at enterprise scale.” Teams running real workloads hit the ceiling: limited multi-agent coordination, no shared memory across agents, weak observability, and a tight coupling to Bedrock’s own model roster.

What Is AgentCore?

AgentCore is AWS’s answer to those limitations. It’s described as framework-agnostic infrastructure for building, deploying, and operating AI agents — not a managed agent service, but the plumbing underneath one.

The key capabilities:

• Runtime — containerized, auto-scaling execution for any agent framework (Strands, LangGraph, CrewAI, or custom code)
• Identity — dedicated Agent Identity credentials that follow least-privilege access and log every operation
• Memory — shared persistent memory across agents in a multi-agent system
• Gateway — a unified routing layer that normalizes tool calls and manages external API integrations
• Browser — a managed browser instance agents can operate headlessly
• Code Interpreter — sandboxed code execution for data and analysis tasks
• Observability — full tracing and logging built in, not bolted on

The architecture shift is significant. Classic was one opinionated managed service. AgentCore is a suite of modular infrastructure primitives you compose yourself, or plug into an existing framework.

Who Is Affected?

New teams on AWS: You cannot start with Bedrock Agents Classic anymore. AgentCore is the path forward for any new agent development.

Existing Classic users: Your workloads are unaffected for now. No end-of-life date has been announced. But the model catalog is frozen, meaning any new foundation models AWS adds to Bedrock after July 30 are only available through AgentCore. If you’re using agents for tasks where model quality matters and you expect to upgrade models over time, that’s a meaningful constraint.

Teams mid-migration: AWS provides an AgentCore Import-Agent toolkit that converts existing Bedrock Agents Classic configurations to LangGraph. AWS estimates the migration timeline at two to four weeks for most workloads.

What This Means for Business

The shift from Classic to AgentCore reflects a broader maturity curve happening across the AI agent industry. The first generation of agent tooling — where you configure a system and a managed service handles the rest — turns out not to scale well when you need multiple agents working together, sharing memory, and operating within strict security and compliance guardrails.

AgentCore is AWS’s acknowledgment that enterprise AI agents are infrastructure, not just software. The plumbing needs to be solid, observable, and composable — the same principles that govern any serious cloud workload.

For businesses evaluating AI agent infrastructure right now, a few practical implications:

Vendor lock-in just got a bit looser. AgentCore’s framework-agnostic approach means you can run LangGraph agents on AWS without being tied to Bedrock’s model catalog. You bring the framework; AWS provides the runtime, memory, and gateway. That’s a more defensible architecture than Classic’s tight coupling.

Multi-agent systems are now first-class. Classic was designed for single agents doing sequential tasks. AgentCore was designed from the start for systems of agents that delegate, collaborate, and share state. If you’re building anything beyond a simple chatbot or single-task automator, this distinction matters.

Observability is table stakes, not optional. The fact that AgentCore ships with built-in tracing and logging — rather than asking teams to wire it up themselves — signals what serious enterprise deployments require. If you’re running agents in production and you can’t answer “what did that agent actually do?”, you have a governance problem.

The three-year lifecycle is a reality check. Bedrock Agents launched in November 2023 and reached end-of-life for new customers by July 2026. That’s a fast cycle for enterprise infrastructure. Teams building AI agent systems today should evaluate whether the platforms they choose have credible long-term architectures — or whether they’re making the same temporary bet on tooling that will need a migration in two years.

What This Means for Business

If your team is on AWS and using — or evaluating — Bedrock Agents, the decision tree is straightforward:

• New projects: Start with AgentCore. There’s no reason to build on a frozen platform.
• Existing workloads: Assess your model refresh cadence. If you’ll need newer models in the next 12 months, start migration planning now.
• No timeline pressure yet: AWS has not announced when Classic will be shut down for existing customers. But maintenance mode means you’re on borrowed time — plan accordingly.

Enterprise DNA works with businesses across industries that are building or evaluating AI agent infrastructure. The shift from Classic to AgentCore is exactly the kind of architectural decision that looks straightforward on the surface but has real implications for long-term maintainability and cost. If your team is navigating this, connect with our advisory team to think through the right approach for your specific workloads.