Why this is the AWS topic to watch right now
Ask anyone tracking AWS in 2026 what the platform's biggest shift has been, and the answer is consistent: agentic AI has moved from demo to default. AWS has spent the year rapidly building out Amazon Bedrock AgentCore, a purpose-built platform for building, deploying, and operating AI agents at enterprise scale — and the pace of updates (new features nearly every month, most recently a managed consent portal shipped in the first week of September) makes it one of the most actively evolving corners of the entire AWS ecosystem right now.
If you've heard AWS talk about "agentic AI," "AI agents in production," or seen the retirement notice for the original Bedrock Agents feature, AgentCore is the reason why.
What Amazon Bedrock AgentCore actually is
AgentCore is best understood as infrastructure for agents, rather than a single agent product. It's framework-agnostic and model-agnostic: teams can bring agents built with Strands (AWS's own open-source agent SDK), LangGraph, CrewAI, or another toolkit, and run them on AgentCore's managed runtime rather than stitching together their own hosting, security, and observability layers.
The platform is composable, made up of several services that can be used together or independently:
- Runtime — hosts agents with full session isolation, so each conversation or task runs in its own protected environment, and supports long-running workloads that don't fit the request/response pattern of typical APIs.
- Gateway — manages how agents discover and call tools and external APIs, effectively acting as a permissioned tool registry.
- Memory — gives agents persistence across interactions, supporting different memory types (things like semantic recall, episodic history, and procedural know-how) so an agent can build context over time instead of starting fresh every session.
- Identity — handles how agents authenticate and act on a user's behalf, using OAuth 2.0 and OpenID Connect flows.
- Code Interpreter — a sandboxed execution environment agents can use to run code as part of completing a task.
What's new: a fast-moving 2026 timeline
A few milestones show just how quickly AWS has been iterating:
- April 2026 — AWS added a managed harness (in preview) that lets a developer define an agent with just a model, a system prompt, and a set of tools, and run it immediately without writing orchestration code. Each session spins up its own isolated microVM, and developers can later export the underlying orchestration as Strands-based code for more control. An AgentCore CLI arrived alongside it, along with built-in "skills" that give coding assistants accurate, up-to-date AgentCore guidance.
- April–May 2026 — AgentCore added Payments in preview, developed with Coinbase and Stripe, which lets agents autonomously access and pay for APIs, tools, and even other agents — handling wallet authentication, transaction execution, and spending controls without custom billing integrations.
- Spring 2026 — an evaluation and optimization loop rolled out, including recommendations generated from production traces, batch evaluations, and A/B testing, so teams can systematically improve agent quality instead of relying on manual tuning.
- May 2026 — AgentCore launched in AWS GovCloud (US-West), extending agentic AI to workloads with elevated compliance requirements.
- August 2026 — AgentCore expanded into two more regions, US West (N. California) and Asia Pacific (Hyderabad), bringing lower latency for agents serving customers in those areas.
- September 2026 — AgentCore Identity introduced a managed consent portal, removing the need for developers to build and host their own OAuth callback infrastructure when connecting agents to third-party services like GitHub, Salesforce, or Slack. Each Gateway now gets its own hosted consent page that admins can share with their team before an agent starts acting on their behalf.
Why AWS is pushing so hard on this
The clearest signal of how central AgentCore has become is that AWS is retiring its predecessor. The original Amazon Bedrock Agents feature — launched back in November 2023 — was rebranded "Bedrock Agents Classic" and stopped accepting new customers as of July 30, 2026. Existing users can keep running on it, but AWS's official guidance is to migrate to AgentCore, which tells you where the roadmap and investment are headed.
AgentCore also sits inside a bigger agentic push across AWS this year: the open-source Strands Agents SDK for building agents, the consumer-facing Amazon Quick assistant (with its own desktop app), and an expanded partnership with OpenAI that's bringing models like GPT-5.5 to Bedrock along with "Managed Agents powered by OpenAI," currently in limited preview. Put together, it's a coordinated bet that the next phase of cloud computing competition will be won on who can host, secure, and scale AI agents most reliably — not just who has the best foundation model.
What it means if you're building on AWS
For engineering teams, the practical takeaway is that a lot of the undifferentiated heavy lifting around agents — session isolation, tool permissioning, memory, identity, payments, and evaluation — is becoming something you configure rather than something you build from scratch. That's meaningful if your organization has been prototyping agents with a single framework and hitting a wall on security review, observability, or getting from a working demo to something a compliance team will sign off on.
It's also worth watching if you're on Bedrock Agents Classic today: with new customer sign-ups already closed, planning a migration path to AgentCore before it becomes urgent is the sensible move.
Agentic AI has been the buzzword of the year across every cloud provider, but AWS's approach with AgentCore — a modular, framework-agnostic platform rather than a single walled-garden agent product — is arguably the most concrete expression yet of what "production-grade agents" is supposed to mean in practice.
# agents# amazon-bedrock# amazon-bedrock-agents# amazon-bedrock-agentcore# knowledge-bases-for-bedrock
Any opinions in this article are those of the individual author and may not reflect the opinions of AWS.Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article