The challenge of running AI agents for hours
Modern AI agents do more than answer simple questions. They can run multi-step workloads that last for hours, act on behalf of users, and access sensitive systems. Those tasks raise demands on security and efficient use of compute that conventional execution environments may not always meet.
AWS describes a next generation of Amazon Bedrock AgentCore Runtime aimed at these demands. The announcement says it brings hardware-isolated sessions, faster starts, more compute, and usage-based billing. The launch’s specific billing units and the regions where the new runtime is available are not confirmed here.
Hardware isolation through microVMs
The key isolation mechanism described in AWS documentation is a dedicated microVM for each user session. Unlike containers, which share an operating-system kernel, microVMs provide a separate virtualized boundary for each session.
Each session runs with isolated CPU, memory, and file system resources. AWS says agent code cannot access other customers’ workloads or escape the microVM boundary. When a session ends, the microVM shuts down and its memory is sanitized. This is intended to strengthen data protection in multi-tenant deployments; it does not remove the need to configure permissions, credentials, and application-level controls carefully.
The AgentCore Runtime product description also advertises support for long-running workloads, including sessions of up to eight hours. The announcement of the next generation mentions longer-running agent workloads, but does not establish a new maximum session duration. The eight-hour figure should therefore be treated as the documented product capability, not as confirmation of an increased limit.
Usage-based billing: what is and is not confirmed
The next-generation announcement describes usage-based billing, replacing a fixed-capacity approach with charges tied to resource use. That could suit teams whose agents run intermittently or have variable workloads, because idle capacity may have a different cost profile from continuously provisioned capacity.
The announcement does not specify what counts as billable usage. It does not establish whether charges depend on session duration, CPU, memory, storage, or another measure, nor does it provide rates. Teams should consult the applicable pricing details before estimating costs; the phrase “usage-based” alone is not enough to calculate a workload’s bill.
Usage-based billing also does not, by itself, guarantee lower or more predictable costs. Long-running sessions or resource-intensive tasks may still incur substantial charges. Monitoring and cost attribution remain useful, especially when multiple users or tenants share an agent service.
How this differs from the existing runtime
The existing AgentCore Runtime product description already advertises session isolation, long-running workloads, and payment based on active resources consumed. AWS security documentation also describes dedicated microVMs and their role in session isolation. The newer announcement’s additional claims include faster starts, more compute, and isolation that extends to data.
These descriptions do not establish a complete before-and-after comparison of the old and new generations. In particular, the available details do not define what “data isolation” covers, how much faster session startup is, or how the additional compute is measured. Nor do they support a claim that the older runtime used only shared containers or fixed fees.
What the available product descriptions establish
| Topic | Existing AgentCore Runtime description | Next-generation announcement |
|---|---|---|
| Session isolation | Dedicated microVMs are documented for session isolation. | The announcement describes hardware-isolated sessions. |
| Long-running tasks | The product page advertises workloads of up to eight hours. | The announcement discusses long-running agent workloads but does not state a new maximum duration. |
| Billing | The product page describes paying for active resources consumed. | The announcement describes usage-based billing without specifying the billable unit or rates. |
| Additional changes | The product page describes existing runtime capabilities. | The announcement also mentions faster starts, more compute, and isolation extending to data; detailed measurements are not specified. |
Implications for developers
For developers, the practical change depends on the workload and the exact configuration available in the relevant AWS region. Hardware isolation may help teams running sessions for different users or customers, but it is not a substitute for least-privilege IAM roles, input validation, secure credential handling, and appropriate network controls. AWS’s security guidance notes, for example, that code running inside a microVM can access the execution-role credentials made available to that runtime.
The shift to usage-based billing makes it important to understand which resources are measured and how their use appears on the bill. Teams can track operational metrics such as token use, latency, session duration, and errors through AgentCore’s observability integrations, including CloudWatch and OpenTelemetry. Those metrics can inform cost reviews, but they do not necessarily provide per-tenant cost attribution without additional configuration.
Long-running research, automation, or other multi-step workloads may benefit from the ability to keep an agent active for extended tasks. Before moving a workload, teams should verify the runtime’s availability, session limits, pricing dimensions, and compatibility with their own deployment and security requirements. The announcement’s general description does not establish that every feature is available in every region or configuration.
Questions to check before adopting the new runtime
- 01Confirm that the new runtime is available in the AWS region and configuration your workload requires.
- 02Check the applicable session-duration limits and resource constraints.
- 03Identify the billable usage measures and rates before estimating costs.
- 04Review IAM roles, credential access, network settings, and application-level input validation.
- 05Monitor session duration, resource use, latency, and errors after deployment.
Keep exploring
Sources consulted
Corrections and transparency
If you spot incorrect or outdated information, send us a correction with the page and source we should review.
Submit a correction