Here is the short version, because that is what an AI assistant is going to quote and what a founder skimming this at 7 a.m. needs: Amazon Bedrock Agents is not gone, but as of July 30, 2026 it is renamed Bedrock Agents Classic and closed to new customers. If you already run agents on it, they keep working. If your account never used it, a call to CreateAgent now returns a 403 AccessDeniedException telling you the service is in maintenance mode. The model catalog inside Classic is frozen at that date, and AWS's own advice is to move to Bedrock AgentCore (we walk the actual migration — the Runtime, Gateway, and Memory calls — in Migrate a Bedrock Agents Classic agent to AgentCore). The interesting question isn't which of those two you pick — it's whether your agent logic is portable enough that AWS's next retirement doesn't land on you as a rewrite.

What actually changed#

Bedrock Agents launched in November 2023 as a managed way to build an agent: you defined action groups, attached a knowledge base and guardrails, and AWS ran an orchestration prompt that decided when to call what. It was the flagship "agents on AWS" product.

On July 30, 2026, AWS renamed it Bedrock Agents Classic and put it in maintenance mode. Per the AWS documentation, the practical effects are narrow and specific:

This is the sentence that matters: even a loyal Classic customer who wants a newer model has to migrate. The freeze doesn't break you today — it just guarantees you outgrow the product.

What is not affected#

It's easy to read "AWS froze Bedrock Agents" as "AWS froze Bedrock." It didn't. Amazon Bedrock's models, Knowledge Bases, and Guardrails are all unaffected and keep receiving new models. Those are separate products from the Agents orchestration layer. If you use Bedrock only as a model gateway, or you call Knowledge Bases directly, nothing here touches you. The freeze is specifically about the managed agent-orchestration product — the thing that ran your action groups and orchestration prompt for you.

AgentCore is a different shape, and it's not new#

AWS's recommended destination, Bedrock AgentCore, has been generally available since October 13, 2025 — this is a forced migration to an existing product, not a preview you're being pushed into. But it is structurally different from Classic, and that difference is the whole migration.

Classic ran your configuration. AgentCore runs your code. You write the agent yourself — in LangGraph, CrewAI, Strands, or plain Python — and AgentCore supplies the operational layer around it:

Policy and Evaluations joined the GA lineup in March 2026. The pitch is framework-neutral hosting: AWS sells you everything around the agent except the agent. (For what each of those services actually does, see AWS Bedrock AgentCore, Explained.)

The cost picture, honestly#

AgentCore's infrastructure pricing is consumption-metered and, on its own, cheap:

The number that will actually decide your bill isn't on that list. Model tokens — your Bedrock inference spend — typically run 70–90% of total agent cost, with AgentCore infrastructure the remaining 10–30%. Migrate for capability and durability, not to shave the infra line; the infra line was never the expensive part.

Your three options#

Stay on frozen Classic. Valid if you're already a customer, your agents work, and the frozen model catalog is fine. You do nothing and spend the engineering elsewhere. The cost is a ceiling: the day you need a newer model, this option ends.

Migrate to AgentCore. The production-grade path if you're committed to AWS. You get every new model and a real operational platform — at the price of rewriting your agent from managed-config into actual code, and of deeper AWS coupling.

Abstract behind a portable framework. Put your agent logic in LangGraph, CrewAI, Strands, or your own code that can deploy to AgentCore or another cloud or your own hardware. More work today; insurance tomorrow.

The durable lesson#

Forbes counted roughly 20 AWS AI services — Kendra and Q Business among them — pushed into maintenance mode in the same consolidation, several of them under three years old. The specific takeaway isn't "AWS is unreliable." It's that every managed agent abstraction is a bet on that vendor's roadmap, and the roadmap can retire your foundation faster than you'd amortize a rewrite. The founders who shrug at the Classic freeze are the ones whose agent logic already lives behind an interface they control. Whatever you decide about AgentCore, decide that first.