SeaVerse, a gaming startup from SeaArt, builds a creation-first platform for playable AI experiences: users can open lightweight games, chat with AI characters, run interactive apps, or create their own experience from a prompt. Every creation passes through the same execution chain — generate, run, preview, debug, publish, remix — and any latency, instability, or isolation failure in that chain is immediately felt by users. To support this, SeaVerse needed dynamic multi-tenant sandbox workloads with strong security boundaries, low latency, better observability, and more flexible cost.

Existing sandbox approaches did not offer enough kernel-level isolation or native observability, so diagnosing failures in multi-tenant environments was slow and costly. SeaVerse chose Google Kubernetes Engine (GKE) and GKE Agent Sandbox to avoid turning its team into a cluster-maintenance group while keeping the Kubernetes ecosystem and operational tooling. GKE Agent Sandbox is a Kubernetes primitive designed for agent code execution and computer use. With Kata Containers plus Cloud Hypervisor (microVM), it provides strong security boundaries without slowing the creation experience, and teams can switch between microVM and gVisor isolation runtimes. Built on gVisor, it gives kernel-level isolation for dynamic sandbox workloads while preserving the same Kubernetes scheduling, monitoring, and operations as the rest of the cluster.

At general availability, GKE Agent Sandbox supports allocating up to 300 sandboxes per second per cluster, with 90% of allocations completing within 200 milliseconds. Native logging and monitoring reach directly into the sandboxed environments, shifting SeaVerse from a black-box debugging experience to a glass-box one. The team can now see runtime status, metrics, and failure signals, which speeds up issue resolution and strengthens management of multi-tenant workloads. The isolation model also changed cost planning: instead of depending on specific server types, SeaVerse can run secure workloads on appropriately sized cloud VMs, which helped cut infrastructure costs by up to 60%. The platform also gained persistent-storage support for creations that evolve, while preserving isolation boundaries.

Looking ahead, SeaVerse plans to scale toward over a million sandboxes and continue improving stability, observability, elastic scaling, and cost efficiency. It is also exploring additional Google Cloud tools: Gemini and agent models for helping operators and creators understand experience performance, BigQuery AI and ML capabilities for use cases such as churn prediction, LTV and ROI prediction, and user segmentation, plus multimodal tools such as Imagen and Veo on Gemini Enterprise Agent Platform for creative generation and AI interactive content production.