Watch Abubakar Siddiq Ango's talk
AI agents that interact with Kubernetes face a boundary problem. A shared cluster exposes its entire API surface, and there is no way to scope what an agent can discover or act on for each tenant. kcp, a CNCF Sandbox project, solves this with workspace-level isolation. Each workspace is a standalone Kubernetes API server with its own scoped OpenAPI surface, and none of them needs a dedicated cluster. In this session, Abubakar Siddiq Ango connects Google’s open-source Agent Development Kit (ADK) to kcp workspaces. ADK’s OpenAPI toolset discovers available operations automatically from kcp’s API specs, so an agent can provision and manage workloads across isolated workspaces using natural language. The demo shows how kcp’s API export model lets agents compose platform services across team boundaries, while API-driven syncing connects workspaces to physical service clusters. The result is a working loop from natural language to running workloads.
Key Takeaways
- Scoped Agent Access: How kcp workspaces give each tenant an isolated API surface, which limits what AI agents can see and do.
- Natural Language to Workloads: How ADK’s OpenAPI toolset auto-discovers kcp operations to provision and manage workloads through natural language.
- Composable Platform Services: How kcp’s API export model and API-driven syncing let agents work across team boundaries and reach physical service clusters.
