What Is Bedrock Agent Runtime? The API vs AgentCore Runtime

Two different things share the name

"Bedrock agent runtime" points at two AWS products that do different jobs. Mixing them up is the most common reason the documentation seems to contradict itself.

Agents for Amazon Bedrock Runtime is an API. It is the runtime half of the original Amazon Bedrock Agents service, the endpoint your application calls once an agent exists. In the AWS SDKs it is the bedrock-agent-runtime client, separate from the bedrock-agent client used to create and configure agents.

Amazon Bedrock AgentCore Runtime is a hosting service. You package your own agent code as a container and AgentCore runs it, one isolated session at a time. Its operations live in a different namespace: bedrock-agentcore-control to create runtimes, bedrock-agentcore to invoke them.

The Agents for Amazon Bedrock Runtime API

AWS's API reference lists the operations under Agents for Amazon Bedrock Runtime. They fall into four groups, and only one of them is about agents in the narrow sense.

Agent invocation. InvokeAgent sends input to a configured agent alias and streams back the response and trace. InvokeInlineAgent runs an agent defined in the request itself. Session and memory operations, such as CreateSession and GetAgentMemory, sit alongside.

Knowledge base retrieval. Retrieve, RetrieveAndGenerate, its streaming variant, and Rerank query Amazon Bedrock Knowledge Bases. Many developers who type this phrase are building retrieval, not agents, and this is where those calls live.

Flows. InvokeFlow and the flow execution operations run Bedrock Flows.

Utilities. OptimizePrompt and GenerateQuery, among others.

In this model AWS owns the agent loop. You declare a model, instructions, action groups, and knowledge bases; Bedrock orchestrates the steps.

The status change that matters

AWS has renamed Amazon Bedrock Agents, launched in November 2023, to Amazon Bedrock Agents Classic and put it in maintenance mode. According to AWS's maintenance mode guide, it is closed to new customers from July 30, 2026. Accounts with Bedrock Agents activity in the previous 12 months are allowlisted and keep full access. Other accounts get an AccessDeniedException when they call CreateAgent or InvokeInlineAgent.

Existing agents keep working, there is no migration deadline, and AWS states no planned end-of-life date. But no new features are planned, and the model catalog available to Agents Classic is frozen as of the July 30 date. Knowledge Bases, Guardrails, and model inference are not affected, so the retrieval operations above remain a normal part of Bedrock.

Amazon Bedrock AgentCore Runtime

AgentCore Runtime is where AWS points new agent work. AWS's developer guide describes it as handling scaling, session management, security isolation, and infrastructure. The details that distinguish it:

Bring your own loop. It hosts containerised agent or tool code built with any framework, and AWS names Strands Agents, LangGraph, and CrewAI among them. Each runtime is versioned, and a DEFAULT endpoint tracks the latest version.

Session isolation. Each session runs in a dedicated microVM with isolated CPU, memory, and filesystem. A session ends after 15 minutes of inactivity or 8 hours of total runtime, and the microVM is then terminated and its memory sanitised.

Protocols. HTTP, MCP for tool servers, and A2A for agent-to-agent communication, plus WebSocket streaming.

Separate services for the rest. Session state is ephemeral by design. Durable context goes in AgentCore Memory, and credentials for inbound and outbound calls are handled by AgentCore Identity.

For teams that liked the declarative style of Agents Classic, AWS also offers a managed harness on AgentCore: declare the model, tools, and instructions, and AgentCore runs the orchestration loop.

Which one people usually mean

Read the context. If the question involves boto3.client("bedrock-agent-runtime"), InvokeAgent, or RetrieveAndGenerate, it is the API, and for retrieval it is the right tool today. If it involves deploying a container, session microVMs, or InvokeAgentRuntime, it is AgentCore Runtime. For any new agent, start with AgentCore; Agents Classic is open only to accounts that already use it.

Whichever you run, the runtime decides where the loop executes, not what the agent knows. Keeping reference material in a retrieval layer such as RDK, instead of in every prompt, cuts token spend 80 to 90 percent on repeated or reference-heavy work.