‘Agents’ Is Closed to Newcomers—Continue with ‘AgentCore’: How AWS Rebuilt the Foundation for Agents in Two and a Half Years
AWS has stopped accepting new customers for its flagship "Amazon Bedrock Agents," appending the "Classic" label and shifting its focus to the successor, "AgentCore." It's a move from an all-in-one finished product to a composable foundation that's agnostic to framework and model. We walk through the
AWS has quietly demoted what was once its flagship service for generative AI agents, "Amazon Bedrock Agents." The official documentation now opens with this disclaimer: "Amazon Bedrock Agents (now Amazon Bedrock Agents Classic) is no longer accepting new customers. For equivalent functionality, please consider Amazon Bedrock AgentCore. Existing customers can continue to use it as before." Launched in 2023 as the standard entry point for building agents on AWS, the service has been given the "Classic" label and has ceded its place to the next generation.
What the Single Word "Classic" Set in Motion
It is not just the name that has changed. According to technical write-ups, the halt on new sign-ups took effect on July 30, 2026; existing deployments enter "maintenance mode," and no new features will be added. Workloads already running in production won't suddenly stop, and you can keep using them for the time being. For anyone building a new agent from here on, however, the entry point is effectively AgentCore and nothing else.
Bedrock Agents was an "all-in-one, finished product" that connected foundation models, data sources, and internal APIs, invoking the necessary processing from a user's conversation. In exchange for AWS handling everything—prompt design, memory, monitoring, and permissions—the flexibility of its configuration was confined within that framework. This line in the sand reads as a statement of intent: to draw a close to that monolithic model and shift the focus to an approach where components are recombined.
From a Standalone "Finished Product" to a Recombinable Foundation
Its successor, AgentCore, describes itself as "a platform for production agents—any framework, any model, secure at scale." LangChain, the OpenAI Agents SDK, the Claude Agent SDK, the Strands SDK, or your own homegrown framework—an agent written in any of these can run on it, and you aren't locked into a particular model. This is a different design philosophy from Classic, which was closed around a specific way of doing things.
AgentCore is not a monolith; it is offered as a collection of components, each with its own role. The main building blocks, as gleaned from the documentation and various write-ups, are as follows.
| Component | Role |
|---|---|
| Runtime | An execution environment that takes the agent code you've built all the way to production and runs it |
| Gateway | A single point that bundles connectivity, authentication, and access control for MCP servers, knowledge bases, internal APIs, Lambda, and more |
| Memory | Persistent memory that retains context across sessions |
| Identity | Permission management that enforces access policies at the platform layer and verifies them through automated reasoning |
| Observability | Observability features that make visible which steps an agent took, what it called, and where it stumbled |
| Evaluation | An evaluation mechanism that tests configuration differences against real traffic and measures quality |
Where Classic assumed a single agent, AgentCore is built on the foundation of orchestrating multiple agents working together. Rather than loading everything onto one agent, you split execution, connectivity, memory, permissions, and observability into independent layers and assemble only the parts you need. It's AWS falling into step with the recent trend in agent design, which has been shifting from "all-in-one" toward a "composable foundation."
The Homework of Migration, and a Warning About the "3x Bill"
Since existing users can keep using the service as-is, there's no need to panic for now. But when the time comes to move to AgentCore, it will involve more than a change of name. One technical write-up broadly organizes the migration paths into three.
- Convert with the toolkit (roughly 2–4 weeks): The shortest route—use the AWS Import-Agent tool to convert to LangGraph format.
- Rebuild (roughly 4–8 weeks): Rebuild from scratch to fit AgentCore's native capabilities.
- Run in parallel: Keep Classic running while gradually adopting AgentCore, starting with the new features.
Whichever path you take, the same write-up notes that Classic's tool configuration needs to be rewritten for the Gateway layer, memory state doesn't carry over automatically, and permissions (IAM) have to be reconfigured to fit a multi-agent setup.
There are cost caveats, too. The same write-up sounds an alarm: "A poorly designed multi-agent configuration can inflate the AgentCore bill to three times that of a standalone Bedrock Agents." That's because requests routed through the Gateway and calls for tool integration now carry usage-based charges that didn't exist before, so carelessly adding components can send the bill soaring. The figure is one estimate rather than an official AWS view, but the direction—"now that it's more flexible, the quality of your design shows up directly on the bill"—is worth keeping in mind.
What This Line in the Sand Tells Us
More than the individual migration steps, the direction is what's worth watching. A major cloud provider has downgraded its own first-generation agent product to "Classic" and shifted its focus to an execution foundation bound to neither a framework nor a model—a move that signals the value of agents is shifting from "a specific vendor's finished product" to "a foundation you can build on anywhere and run." For developers, it's a welcome chance to broaden their options, but it also means greater responsibility for designing yourself how execution, connectivity, permissions, and observability fit together—and how far the charges pile on. A good first step is to take inventory of which entry points your existing agents depend on, so that when the time to decide on migration arrives, you won't be caught off guard.
References: Amazon Bedrock User Guide (the Agents Classic notice) / Amazon Bedrock AgentCore (official) / AWS News Blog: AgentCore adds quality evaluations and policy controls / ServerGurus: What AgentCore Means for Your AI Workloads / AWS in Plain English: AWS Just Retired Its Flagship AI Agent Product
