AWS is adding a persistent compute layer for production AI agents, turning Amazon Bedrock AgentCore into more than a short-running invocation environment.

AWS announced runtime instances for AgentCore Runtime as managed EC2 infrastructure for workflows that need shared state, multiple collaborating agents or GPU access.

The launch addresses a gap between agent prototypes and operational deployments.

AgentCore runtime microVMs already support managed invocations that can run for up to 8 hours, while the new runtime instances are built for jobs that may continue for days or need larger, dedicated environments.

AWS says shared sessions on the same host can persist for up to 14 days.

Persistent Agents Move Into Cloud Infrastructure

Runtime instances let developers deploy multiple agents in a single runtime, with each agent keeping its own dependencies and artifact types.

That structure matters when a workflow needs one agent to write output, another to inspect it and a third to act on the result without rebuilding file transfer or coordination logic around every step.

The official launch material frames the service as a managed substitute for infrastructure that teams previously assembled themselves.

For agents that stay active, use accelerators or work together across one job, customers otherwise needed their own EC2 layer, network setup, scaling path, session logic and monitoring stack.

Runtime instances keep those pieces inside the AgentCore operating model.

The compute choice is not replacing microVMs.

AWS describes runtime microVMs and runtime instances as complementary options through the same AgentCore Runtime APIs.

A lighter orchestrator on microVMs can route tasks and aggregate results, while worker agents on instances handle stateful build jobs, defensive code checks or interface automation when those jobs need direct access to the host operating system.

Runtime Details Define The Deployment Boundary

The specification makes the product a cloud-infrastructure story as much as an AI tooling update.

The launch lists Linux on ARM64 and x86_64, Python 3.11-14 with native code support, container images and GPU-accelerated instance types.

The same AgentCore APIs, identity controls, observability and policy controls apply across the environment.

Pricing keeps the service close to infrastructure planning.

AWS attaches normal EC2 charges to runtime instances and adds an AgentCore orchestration management charge, so customers still need to size long-running agent workloads against compute, storage and idle-time behavior.

Session stop and restart is meant to reduce idle costs when workflows do not need to run continuously.

Availability starts in major AWS regions: the Ohio and Northern Virginia US East regions, Oregon in the western United States, Mumbai, Singapore, Sydney and Tokyo in Asia Pacific, plus Frankfurt and Ireland in Europe.

For enterprise AI teams, the immediate technical test is whether persistent sessions and managed coordination reduce the custom platform work that has slowed agent systems after the prototype stage.