Amazon Bedrock AgentCore Runtime Instances [GA -- August 6, 2026]

A few weeks back we talked about cloud providers racing to make agents a first-class part of the infrastructure layer instead of just an app on top. This week's update is a concrete example landing on AWS specifically.

If you've built anything on Bedrock AgentCore, you know the catch with the default runtime: it's serverless, it's microVM-based, and it caps out at 8 hours per session. Fine for a chatbot. Not fine for an agent doing a multi-day data migration, a long-running compliance sweep, or anything that needs to stay warm and keep context for a while.

AWS just fixed that. Runtime instances, generally available as of August 6, let you run AgentCore agents on your own EC2 instances instead of the managed microVM option. Sessions can now persist for up to 14 days, and you get access to the full range of EC2 instance families, including GPU-accelerated, memory-optimized, and compute-optimized types. AgentCore still handles the provisioning, patching, scaling, and lifecycle work, so you're not back to babysitting infrastructure. You're just no longer boxed into the 8-hour default.

This is a real signal about where AWS thinks agent workloads are headed: less "answer a question and disappear," more "run continuously and do real work." It also puts AWS in more direct competition with what Microsoft and Google have been building. Azure's Foundry Agent Service added hosted agents and persistent memory earlier this summer, and Vertex AI has its own long-running agent runtime. Nobody's settled on one standard architecture for production agents yet, which means the platform you already know is worth watching closely right now.

A few specifics worth knowing about the AWS release:

Multi-agent, shared sessions. You can now deploy multiple agents on a single runtime instance, each with its own dependencies, operating inside sessions that persist across that 14-day window. That's a meaningfully different architecture than spinning up isolated microVMs per request.

GPU access changes the math. If you're running anything that needs a GPU -- embedding generation, local inference, vision tasks -- you couldn't do that on the old serverless runtime. Now you can attach the instance type your workload actually needs.

Regions are limited for now. Runtime instances are live in US East, US West, Asia Pacific, and Europe. If your workload needs to run somewhere else, you're still on the microVM option.

Pricing is standard EC2 plus a management fee. Not free, and not the pay-per-invocation model you get with the serverless runtime. Budget accordingly if you're moving a workload over.

If you're studying for an AWS AI or ML-adjacent cert (AIF-C01, MLA-C01, or one of the newer AI-native tracks) this is worth a spot in your notes even if it's not exam content yet. AgentCore keeps showing up as the answer to "how do I actually run this in production," and cert content tends to catch up to what's shipping about two to three cycles later.

If you're building anything with AgentCore right now, this is the fix for the complaint everyone had about the 8-hour session cap. Worth a weekend to spin up a runtime instance and see what changes in your setup.

And if you're on Azure or GCP instead, don't tune this out. Long-running, persistent-session agents are becoming table stakes across all three platforms. Whichever one you're certifying on, expect this pattern. Agents that run for days, not minutes, to show up in your platform's docs soon, if it hasn't already.

Keep Learning and Building.