Amazon Bedrock AgentCore Guide: Runtime, Sessions, and Agent Endpoints

InYourLeague Research · 2026-08-27

Answer in brief

Amazon Bedrock AgentCore Runtime provides secure, serverless hosting for agents and tools, with framework and model flexibility, isolated sessions, MCP/A2A communication, streaming, authentication, and built-in tracing. Applications can preserve conversation context with session identifiers and process agent output incrementally. The cited documentation does not define a selectable product-level model ID.

Key facts at a glance

Failure modes and verification

FAQ

What does Amazon Bedrock AgentCore Runtime host?

It hosts AI agents or tools in a secure, purpose-built environment. Managed microVMs provide serverless, on-demand sessions, while Instances support persistent sessions, GPU workloads, and collaborating agents on shared infrastructure. See the Runtime overview.

Which frameworks and models can an AgentCore Runtime agent use?

AgentCore Runtime works with LangGraph, Strands, CrewAI, and custom agent code. It can use models available through Amazon Bedrock and models from providers including Anthropic, Google, and OpenAI. See the framework and model guidance.

How are conversation context and session isolation handled?

Use a new, unique runtimeSessionId for each new conversation and reuse it to continue that conversation’s context. In the serverless model, each user session runs in a dedicated microVM with isolated CPU, memory, and filesystem resources. See session management and session isolation.

How do applications invoke an endpoint and stream its output?

Call InvokeAgentRuntime with the endpoint’s runtime ARN, a session identifier, and the expected payload. The operation returns output in streaming chunks, and persistent WebSocket connections support bidirectional streaming. See the invocation guide.

What authentication, protocol, and tracing features are available?

AWS-authorized callers need bedrock-agentcore:InvokeAgentRuntime, and requests can carry a bearer token for user authentication. Runtime supports MCP and A2A communication, AgentCore Identity integrations and outbound authentication, and built-in tracing of reasoning steps, tool calls, and model interactions. See the Runtime capabilities and authentication guidance.

Sources and freshness

• Official source
• Official source
• Last verified: 2026-08-27

Extended guide

Amazon Bedrock AgentCore Runtime is a secure, purpose-built environment for deploying and running agents or tools without managing the underlying session infrastructure. Managed microVMs provide serverless sessions that start and scale on demand, with consumption-based billing. The service also offers Instances on AWS-managed Amazon EC2 infrastructure in the customer’s account for persistent sessions, GPU-accelerated workloads, and multiple collaborating agents on a shared instance.

Amazon Bedrock AgentCore Runtime guide

This guide was verified against the official documentation on 2026-08-27.

Core capabilities

AgentCore Runtime is framework agnostic. It works with LangGraph, Strands, CrewAI, and custom agents that do not depend on a named framework. It can use models available through Amazon Bedrock as well as models from providers including Anthropic, Google, and OpenAI; the documentation names the Claude and Gemini families as examples. Model choice remains part of the hosted agent implementation.

In the serverless execution model, each user session runs in a dedicated microVM with isolated CPU, memory, and filesystem resources. This separation protects stateful reasoning and helps prevent data from crossing between sessions. When a session completes, the entire microVM is terminated and its memory is sanitized. A new runtimeSessionId starts a new conversation, while reusing the same identifier preserves context across requests. microVM sessions can run for up to 8 hours; Instances can support sessions lasting up to 14 days.

Agents can communicate with other agents and tools through Model Context Protocol (MCP) or Agent to Agent (A2A). InvokeAgentRuntime targets a deployed endpoint by its runtime ARN and returns the agent’s output as streaming chunks. It accepts binary payloads up to 100 MB, including payloads for multimodal requests. Persistent WebSocket connections provide bidirectional streaming for interactive applications.

For AWS-authorized invocation, the caller needs bedrock-agentcore:InvokeAgentRuntime. A request can also carry a bearer token for user authentication. OAuth integrations must call InvokeAgentRuntime through HTTPS instead of the AWS SDK. AgentCore Identity assigns distinct identities to agents, integrates with Okta, Microsoft Entra ID, and Amazon Cognito, and supports outbound OAuth or API-key access to services such as Slack, Zoom, and GitHub. Built-in tracing captures reasoning steps, tool invocations, and model interactions for debugging and auditing.

Invoke an endpoint safely

1. Identify the endpoint’s AgentCore Runtime ARN. Add an optional qualifier only when a particular agent version or endpoint must be targeted.
2. Generate a collision-resistant runtimeSessionId, preferably a UUID, for every new conversation. Reuse it only when the conversation should retain its existing context.
3. Encode the payload as the agent entrypoint expects. Because parsed JSON can contain any JSON type, confirm that prompt is a string before passing it to the framework; otherwise, a toolUse block could trigger a tool directly and bypass model reasoning and guardrail evaluation.
4. Inspect contentType and process output as it arrives. Handle text/event-stream, application/json, and any other content types that the deployed agent intentionally returns.
5. Handle errors according to their cause. Correct invalid parameters for ValidationException, verify the ARN and resource for ResourceNotFoundException, and fix permissions for AccessDeniedException. Use exponential-backoff retries for ThrottlingException and short exponential backoff for the transient RetryableConflictException. Use InvokeAgentRuntimeCommand for deterministic shell operations such as tests, Git commands, builds, or environment setup in the same session. Review the available built-in trace data when debugging agent decisions and tool activity.

Model catalog boundary

The cited documentation describes flexible model integration but does not define a selectable model ID for AgentCore Runtime. Do not infer an amazon-bedrock-agentcore model identifier from the product name. See the Runtime overview and invocation guide.

Model availability note: The official source does not specify a selectable model ID.