AX

AX is Google's self-hosted AI agent orchestrator for teams running autonomous workloads on a Kubernetes cluster. It uses Agent Substrate to isolate agent execution and manages the workspaces and model settings each task needs. The project is open source under Apache 2.0.

Its focus is agents that accumulate state and need access to tools while they work. AX runs untrusted agent code in sandboxes with CPU and memory limits. You can pause a task and resume it with its saved state, rather than start the work again.

Workspaces bring together Git repositories, MCP servers and skill packages so agents have their working environment ready. Separate model definitions control which LLM the platform uses, with credentials held in Kubernetes secrets. The orchestration and sandbox execution run in your cluster; agents can call external model APIs and tool servers, so self-hosting AX doesn't make those services local.

For operators, AX provides live task status and access to a running sandbox for inspection. Its command-line interface follows Kubernetes conventions and works with kubectx, which makes it relevant to teams already managing workloads across clusters. Task, workspace and model definitions are declarative, so teams can describe the execution environment alongside the work to perform.

AX also supports custom runner images for teams that need a different agent execution environment.