Amazon Bedrock AgentCore is an agentic platform for building, deploying, and operating highly capable AI agents securely at scale — using any framework (Strands, LangGraph, CrewAI, LlamaIndex…) and any model, with no infrastructure to manage. Its services work together or independently: give an agent the compute to run, the tools and memory to act, and the identity, policy, and observability to do it safely in production.
Managed agent loop · Strands Agents SDK · any framework, model, or harness
Memory · Managed Knowledge Base · Web Search · Browser · Code Interpreter · Payments
Evaluation · Insights · Recommendations · A/B testing
Learn more: AgentCore product page ↗ · documentation ↗
Amazon Bedrock AgentCore Runtime is a serverless runtime purpose-built for running AI agents and tools. You package an agent as a container and AgentCore runs it in an isolated session — its own compute, memory, and filesystem — so it's a safe place to run agentic AI, model-driven tool use, and generated code. It offers two compute types: microVMs (fast, fully serverless, scale to zero) and Instances (EC2 in your own account for longer, heavier, or GPU workloads).
Bring an agent built with any framework (Strands, LangGraph, CrewAI) and any model. If it runs in a container, it runs on AgentCore.
Each runtimeSessionId gets its own dedicated compute and
filesystem, keeping one user's session fully separated from another's.
The first invoke for a session spins up compute; there are no servers or clusters to manage and microVMs scale to zero when idle.
Memory and files persist while a session lives, so an agent can keep context and workspace state between turns — then suspend and resume.
Run microVMs with public egress or inside a VPC; Instances always run in a VPC in your own AWS account.
A single InvokeAgentRuntime call streams the agent's
output back in real time as it works.
You invoke an agent with InvokeAgentRuntime and a
runtimeSessionId. AgentCore ensures compute is running for that
session, forwards your request to the agent, and streams the response back.
The same flow works for both compute types — what differs is the compute
underneath.
The default. Lightweight, API-driven agents that start fast, scale on demand, and finish within hours. Fully AWS-managed, arm64 Linux containers, consumption-based pricing.
AWS-managed EC2 that AgentCore provisions and patches in your account. For long-running, GPU, or multi-agent work — and you keep your own cost controls (Savings Plans, ODCRs).
microVM sessions run up to 8 hours; Instance sessions up to 14 days. Idle sessions suspend to save cost.
When a session stops, invoking the same runtimeSessionId
again resumes it. On Instances, persistent EBS volumes re-attach so the
agent's files are intact.
A microVM runs one agent per runtime (1:1). An Instance session can host multiple agents that share a filesystem and collaborate (1:N).
You choose the compute type when you create the runtime; the invoke API, isolation model, and streaming behavior stay the same either way.
Read the full details in the AgentCore Runtime compute documentation ↗.
Loading scenarios…