Every cloud provider is racing to sell agents a VM. Freestyle.sh , Exe.dev , Vercel, Cloudflare, Google are all converging on the same primitive: spin up a sandbox, let the agent run a script, tear it down. For short-lived sessions, that’s the right primitive. For anything that needs to outlive the session, it breaks.
Agents are being asked to operate as if they’re running locally, but that’s not how production systems work. Developers reach for Docker Compose, Kubernetes, Terraform, and cloud services because a node isn’t enough.
We’ve seen this movie before. In the 2010s, companies outgrew VMs the same way agents are about to. They needed multi-service architectures, custom networking, stateful workloads, horizontal scale, cost control. Kubernetes won because it gave them the primitives of a cloud they could own and operate. Not a box, but a platform.
The base unit shouldn’t be a VM an agent rents. It should be an isolated cloud an agent owns. VMs, storage, networking, identity, DNS, all as composable resources inside that cloud. Basically the agent gets root over a mini-AWS, not SSH into a box.
The obvious move is to give each agent its own Kubernetes cluster. Clusters are the closest existing primitive to “a cloud you own”: namespaces, RBAC, networking, storage, a real API. But the moment you try to operate this at agent scale, every assumption in the stack pushes back.
A cluster takes minutes to provision, not seconds. A control plane costs more per month than most agent sessions will ever generate in value. Multi-tenancy inside a cluster is too soft for untrusted code; one cluster per tenant is too expensive to hand out by the million. The lifecycle is wrong too. Clusters are long-lived infrastructure managed by humans, but agent workloads are bursty, short, and created by software that doesn’t care about upgrade windows.