Amazon Bedrock Agents and AgentCore: running AI agents on AWS

From Agents for Amazon Bedrock to AgentCore

AWS launched Agents for Amazon Bedrock in November 2023: you described an AI agent in the console (instructions, actions, knowledge bases) and AWS handled the orchestration. That service is now called Amazon Bedrock Agents Classic. It is in maintenance mode and has been closed to new customers since July 30, 2026; existing customers can keep using it, but AWS recommends moving to Amazon Bedrock AgentCore, generally available since October 13, 2025.

This is more than a rename. Bedrock Agents Classic offered an AWS-specific agent model; AgentCore takes a different approach: you build your agent however you like, and AWS provides the infrastructure to run it in production.

What AgentCore offers

AgentCore gives you two paths. The first, a managed “harness”, is configured declaratively: model, tools and instructions. The second hosts agents written in code with the framework of your choice, such as Strands Agents, LangGraph, CrewAI, LlamaIndex, the OpenAI Agents SDK, the Claude Agent SDK or Google ADK. AgentCore SDKs are available in Python and TypeScript.

Around the agent, AWS offers separate services: Runtime (isolated execution in microVMs or on EC2), Memory, Gateway (which turns APIs or Lambda functions into MCP tools), Identity, Code Interpreter, Browser, Observability, Evaluations and Payments. AgentCore also supports the A2A protocol, which lets agents talk to each other.

Cost and AWS lock-in

AgentCore pricing is consumption-based with no commitment: runtime is billed per second, based on the CPU and memory actually used. Each service has its own price line, and model costs come on top. As of October 22, 2026, AWS advertises up to $200 in Free Tier credits for new customers. In practice, you will need a measured pilot to estimate costs, as with any pay-as-you-go cloud service.

Lock-in sits at the infrastructure level: identity, runtime, monitoring and billing are tied to your AWS account. It is low on the model and framework side, since your agent code can be written with an open-source tool and reused elsewhere.

Strengths, limits and alternatives

Strengths: a complete production infrastructure, openness to the market’s frameworks and models, and AWS’s security guarantees. Limits: an offering split across many services, a bill that is hard to forecast, and a migration to plan for Bedrock Agents Classic users.

On Google Cloud, the counterpart is Gemini Enterprise Agent Platform. Official maintenance notice: AWS documentation.