Source: AWS News Blog — 2026-08-06
Summary
AWS added "Runtime instances" to Amazon Bedrock AgentCore — a new compute option offering persistent, AWS-managed EC2 infrastructure as an alternative to AgentCore's existing microVM sessions. It's aimed at agents that need to run continuously for days, require GPU access, or run several collaborating agents on a single host, with sessions able to persist for up to 14 days.
Key Takeaways
- AgentCore now offers two distinct compute shapes: short-lived microVM sessions (the original option) and persistent EC2-backed Runtime instances for long-lived workloads.
- Sessions on Runtime instances can persist for up to 14 days — a fundamentally different operating assumption than a microVM that spins up and tears down per session.
- GPU support on Runtime instances opens the door to agents that need local inference or heavier on-host compute, not just orchestration calls to external model APIs.
- Multiple collaborating agents can now share a single host under this model, which changes the deployment story for multiagent systems that previously needed separate isolated sessions per agent.
Reel Script
Hook (16s)
If your AI agent needs to keep working for days, not minutes, a session that resets every time you close the connection is a real problem. AWS just shipped a compute option built specifically for that.
Core Concept (85s)
Bedrock AgentCore's original model runs each agent session inside a microVM — a lightweight, isolated virtual machine that spins up when the session starts and disappears when it ends. That's great for short, stateless interactions, but it's the wrong shape for an agent doing something like a multi-day research task, a long-running monitoring job, or a workflow where several agents need to coordinate continuously. Runtime instances solve that by giving you AWS-managed EC2 infrastructure instead — real, persistent virtual machines that don't tear down between interactions, the way a dedicated desk is different from a hot-desk you have to pack up every evening. The headline number is that sessions on this new compute type can stay alive for up to 14 days, and unlike the microVM option, these instances can be provisioned with GPUs attached, which matters if your agent needs to run local inference or heavier on-host computation rather than just making API calls out to a hosted model.
Hands-On (70s)
The architecture shift worth sketching is: previously, one agent got one microVM per session, torn down at the end. Now you can put multiple collaborating agents on a single persistent Runtime instance, sharing that host across a coordinated multiagent workflow instead of each agent living in its own short-lived box. Picture a diagram with a single EC2-backed instance in the middle, GPU attached, and several agent processes running on it side by side, all able to reference shared state across a session that can legitimately run for two weeks straight — that's a structurally different deployment model than the microVM approach, and it's the piece that actually enables long-horizon or multiagent production workloads on AgentCore.
Takeaway (22s)
If you've been forcing a long-running or multiagent workload into short-lived session infrastructure because that's what your platform offered, this is the option to evaluate — persistent, GPU-capable compute built for exactly that shape of problem. Worth checking whether your current agent architecture is fighting its infrastructure instead of using it.