Deployment Model

The platform runs across two Kubernetes clusters with strict network isolation between platform infrastructure and user workloads.

Two-Cluster Architecture

Cluster	Purpose	IP Space
Trusted	Platform API, platform-admin, guacd (VNC proxy), shared infra	172.32.0.0/16
Untrusted	User agents, Studio containers, customer workloads	172.16.0.0/16

Both clusters are managed via Fleet (GitOps) from the dc_docs/fleet/ directory.

Namespace Model

Each user and team project gets its own Kubernetes namespace in the untrusted cluster:

• Personal projects: one namespace per user
• Team projects: one namespace per project
• Studio, agents, and all user workloads run in the user's/project's namespace

This provides strong isolation — agents from different users can't access each other's resources.

Agent Deployment

Agents deploy as Knative Services, which means they:

• Scale to zero when idle (no cost when not in use)
• Scale up automatically when messages arrive
• Share the same container image as Studio

Config	Value
Image	Studio container image
Entrypoint	/usr/local/xpressai/bin/setup_system.sh continue
NFS PVC	xap-pvc2 (shared across namespaces)
NFS subPath	hdd/user/{userId}/ (personal) or hdd/user/projects/{projectId}/ (team)
Mount	/data/home

Agent files live at /data/home/agents/{agentName}/ inside the container.

NFS Storage Layout

xap-pvc2/
  hdd/user/
    {userId}/                          # Personal project
      agents/
        {agentName}/
          agent.yaml
          tools/
          modules/
          prompts/
          procedures/
          skills/
          knowledge/
      knowledge/                       # User's knowledge base
    projects/
      {projectId}/                     # Team project
        agents/
          {agentName}/
            ...same structure...
        agents/shared/knowledge/       # Shared team knowledge

Domain & Networking

note

The domain names listed below are for the default production deployment. Actual domain names may vary depending on your deployment environment.

• Platform API: platform.ap.xpressai.cloud
• Agent services: {serviceName}.{namespace}.ap.xpressai.cloud
• LLM Relay: relay.public.cloud.xpress.ai/v1 (OpenAI-compatible proxy)

Cold-Start Latency

Because agents deploy as Knative Services that scale to zero when idle, the first request after an idle period incurs cold-start latency. This typically includes container startup, NFS mount, and module initialization. Subsequent requests within the active window are handled immediately. The cold-start time varies depending on agent complexity but is generally a few seconds.