AWS has introduced Runtime Instances, a new Amazon Bedrock AgentCore option for agents that need to work for hours or days. The instances can share resources, use selected GPUs, and maintain sessions for up to 14 days.
Why are persistent instances needed?
Some agents complete their tasks in a few minutes. Others need to research, transform data, run tools, or coordinate several processes over longer periods.
AgentCore microVMs are designed for shorter tasks and sessions of up to eight hours. Runtime Instances provides persistent infrastructure for longer jobs, agent collaboration, and workloads that require GPUs.
Sessions of up to 14 days
A session can be configured to last up to 14 days. This does not mean that every session will have that duration or can remain active indefinitely.
Teams can define a maximum lifetime and close sessions that remain idle. These options help prevent infrastructure from continuing to consume resources after it is no longer needed.
Collaboration among several agents
Several agents can work on the same instance, share files, and divide a complex task.
For example, one agent can collect information, another can analyze it, and a third can prepare the result. Collaboration remains subject to the organization’s permissions, access roles, networking, and controls.
Persistent files and GPU access
The environment can preserve files, installed packages, and results across pause and resume cycles. Amazon EBS volumes can also store information that needs to remain available for longer periods.
Some configurations support GPUs for inference, simulation, rendering, and other intensive tasks. Acceleration is not included automatically with every instance and must be selected when the infrastructure is configured.
AWS manages the infrastructure
The instances run on Amazon EC2 inside the customer’s AWS account. AgentCore manages provisioning, updates, scaling, and machine teardown.
The organization retains control over its data, networks, permissions, storage, and infrastructure type. Runtime Instances also supports different models and agent development tools, so it does not require a single provider.
Availability and costs
Cost depends on the instance type, usage time, storage, and possible GPU use. Pausing processing may reduce consumption, but persistent volumes and other associated resources can continue to generate charges.
AWS does not provide one complete regional list in the announcement. Availability should be checked in the AgentCore console and documentation before configuring an application.
