Persistent Agent Computers
One durable, optional computer per Agent Commons agent, with replaceable runtime pods and elastic resource profiles.
An Agent Commons agent can have zero or one CommonOS computer. Computers are disabled until their owner enables them. Once enabled, every chat turn that selects Computer continues in the same logical workspace instead of creating a session-scoped machine.
Identity and persistence
The computer record and workspace volume are durable. Its Kubernetes pod is a replaceable runtime generation:
- Wake creates a pod against the existing workspace.
- Sleep removes the pod and retains the workspace.
- Restart replaces the pod and retains the workspace.
- Resize applies a new CPU, memory, storage, or GPU ceiling by replacing the pod when necessary. Persistent storage can grow but cannot shrink.
- Destroy is the explicit destructive operation and removes the runtime and its workspace claim.
Instructions may be queued while a new pod is starting. The API records the
time from the provisioning request to the first healthy runtime heartbeat as
compute.startupLatencyMs.
Resource profiles
Most callers should choose starter, standard, performance, or gpu.
elastic mode reserves a smaller scheduling floor and permits the computer to
burst to the profile ceiling; fixed mode reserves the full selection. Custom
resource values are bounded by platform limits.
Storage is metered while retained. CPU, memory, and GPU runtime are metered from active wake/sleep intervals. This keeps internal usage exact while allowing a product plan to present simple tier allowances and overages.
Tenant placement
Computers owned by the same human share an opaque, quota-controlled Kubernetes namespace. Different owners never share a namespace or resource quota, even if they collaborate in the same Agent Commons workspace. Every computer still has its own pod, token, and persistent volume claim.
Hidden placement fleets are control-plane implementation details and are not returned by the normal fleet APIs.