Infrastructure

AWS Extends AI Agent Sessions From 8 Hours to 14 Days

AWS launched persistent, GPU-backed compute for AI agents on August 6, 2026, extending sessions from 8 hours to 14 days and shipping in Frankfurt and Ireland. Why this turns AI compute into a server-budgeting problem.

InfrastructureBy Servola Tech Desk2026-08-133 min read

AI-assisted, edited by humans. | Editor-in-Chief: Leon Soliman | Editorial standards

AWS Extends AI Agent Sessions From 8 Hours to 14 Days

Key takeaways

AWS launched Runtime instances for Amazon Bedrock AgentCore on August 6, 2026, letting AI agents run in persistent, GPU-accelerated EC2 infrastructure for up to 14 days per session, a roughly 40-fold extension of the 8-hour ceiling that applied to the platform's original microVM-based invocations.
The feature is available at launch in ten AWS regions, including two in the EU - Frankfurt and Ireland - alongside US East Ohio and Northern Virginia, US West Oregon, and four Asia Pacific regions, putting long-running agent workloads within EU data residency reach from day one.
Multiple agents deployed to the same runtime instance collaborate by reading and writing files to a shared session directory rather than calling each other through APIs, which is a materially different trust boundary than most agent-to-agent architectures assume and one that security reviewers evaluating the service need to model explicitly.
Pricing combines standard EC2 instance rates with an additional AgentCore orchestration management fee, and sessions can be stopped and restarted to avoid paying for idle time - but the shift from per-call inference billing to provisioned, multi-day compute means AI workloads now show up on a cloud bill the way traditional persistent servers do, not the way stateless API calls do.
The launch quietly answers a question enterprises have been asking since agentic AI moved from demos to production: agents that need to run for days, coordinate through shared state, or use a GPU for sustained periods no longer fit AWS's original serverless agent model, so AWS built a second, server-shaped one alongside it rather than stretching the first.

What Shipped on August 6, and Where

AWS's own announcement describes Runtime instances as 'persistent, managed EC2 infrastructure for production AI agents with multi-agent collaboration, GPU support, and sessions lasting up to 14 days.' That figure is the headline change: the AgentCore Runtime microVMs AWS shipped previously cap out at 8 hours per invocation, which is workable for a single task but not for an agent meant to keep working, monitoring, or waiting on external events across days.

Regional availability at launch spans US East (Ohio, N. Virginia), US West (Oregon), Asia Pacific (Mumbai, Singapore, Sydney, Tokyo), and Europe (Frankfurt, Ireland). For any EU organisation that needs agent workloads to keep data processing inside the EU for residency or contractual reasons, the Frankfurt and Ireland availability at launch - rather than months later - is the detail worth noting.

How Agents Actually Talk to Each Other: A Shared Directory, Not an API

AWS's own worked example is concrete: a code-writing agent saves its output to a path like /tmp/agentcore-session/{session-id}/code.py, and a second reviewing agent reads from that identical path using the same session ID, coordinating through the shared file system a runtime instance provides within a session rather than through direct API calls between the two agents.

That is a meaningfully different architecture from the request-response pattern most agent orchestration frameworks use, and it matters beyond convenience. A shared session directory is a shared trust boundary: any agent with write access to that path can, in principle, alter what another agent reads, which is a lateral-movement surface inside what AWS presents as a single isolated session and one that a security review of AgentCore needs to test explicitly rather than assume away.

The Real Shift: AI Compute Now Shows Up on the Bill Like a Server

Pricing for Runtime instances is standard EC2 instance pricing plus an AgentCore orchestration management fee, and AWS's own example configuration - a c7g.2xlarge instance with 8 vCPUs and 16 GiB of memory, running Arm64 or x86_64 Linux - reads like a server specification because it is one. Sessions can be stopped and restarted to avoid paying for idle periods, but a provisioned instance that persists for up to 14 days is a fundamentally different cost object than a stateless inference call billed per token.

For any finance or platform team that has been budgeting generative AI spend purely as an inference-cost line item, this is the point at which agentic AI workloads start behaving like the rest of the compute estate: capacity planning, idle-time management, and instance-type selection all become live budgeting questions again, the same way they were before serverless abstracted them away.

What an EU Buyer Should Check Before Adopting Runtime Instances

Two questions are worth resolving before a production deployment. First, for any workload processing personal or regulated data, confirm that Frankfurt or Ireland availability actually satisfies the organisation's specific residency and processing-location requirements, since regional availability of the service is not automatically the same as a guarantee about where every piece of session data and log output lands.

Second, for any multi-agent deployment using the shared session directory pattern, ask the same question a security team would ask of any shared-filesystem architecture: which agents can write to the shared path, what happens if one agent's output is malicious or corrupted, and whether the isolation AWS provides between separate sessions is the isolation boundary the organisation actually needs between the individual agents running inside one.

Frequently asked questions

Does this replace the original 8-hour AgentCore Runtime microVMs?

No. AWS is offering Runtime instances alongside the existing microVM-based invocations rather than replacing them; the microVM model remains available for shorter, stateless tasks, while Runtime instances target workloads that specifically need persistence beyond 8 hours, GPU access, or multi-agent coordination through shared state.

Does launching in Frankfurt and Ireland mean agent data automatically stays in the EU?

Regional availability means the compute can run in those regions, but it does not by itself guarantee that every piece of session data, logging, or metadata stays within EU boundaries; an organisation with residency obligations should verify AWS's specific data-flow documentation for Runtime instances rather than infer full data residency from regional service availability alone.

How is Runtime instances pricing different from a standard EC2 instance?

The underlying compute is billed at standard EC2 rates, but AWS adds an AgentCore orchestration management fee on top for the session management, GPU acceleration support, and shared-filesystem coordination features specific to the service; AWS has not published that fee rate publicly as a fixed figure separate from the EC2 base cost.

AWS did not make its serverless agent runtime last longer - it built a second, server-shaped runtime next to it, because 14 days of persistent, GPU-backed compute with a shared filesystem is not a stretched serverless function, it is a server. Buyers who keep budgeting and reviewing it as if it were the former will miss both the cost line item and the trust-boundary question the latter actually requires.

AWSAmazon BedrockAgentCoreAI agentscloud computeEU data residency

Autonomous AI Agents Attacked Taiwan's Government

Over four days in July 2026, up to eight AI agents built on open-source tools breached 85+ Taiwan government accounts and reached the grid. Why this is an EU risk too.

4 min read

Artificial Intelligence

Nvidia's Open-Source Router Cut Ramp's AI Bill 58%

Nvidia's new NeMo Switchyard router already cuts Ramp's coding-agent costs 58 percent and LangChain's 74 percent by calling frontier models only rarely. What the shift means for AI budgets.

7 min read

Decisions

The Third Pivot Was the Warning Magic Leap Missed

Magic Leap filed a WARN notice for 193 layoffs effective 1 October 2026, its fourth strategic pivot after $4 billion raised. The lesson is the pattern, not the pivot.