Google open-sources AX, a declarative orchestrator for agentic tasks

Google has open-sourced AX, a declarative orchestrator for agentic tasks. The idea is the same as in Kubernetes: you describe the desired state in manifests, and the system brings the cluster to match it.

Everything is expressed as ax.io/v1alpha1 manifests and applied with a single command such as ax apply -f task.yaml. A single file can describe a task, a working environment, a network gateway and a model all at once.

There are four basic primitives. Task — an isolated sandbox with CPU and memory limits where untrusted code is executed. Workspace — pre-connected git repositories, MCP servers and skill packages so the agent starts off "warm". Gateway — outbound traffic is trimmed to an explicit list of allowed hosts. Model — which LLM runs under the platform's hood, with credentials taken from a Kubernetes secret.

Control is deliberately built in the kubectl style: apply, get, describe, watch, delete plus a few agent commands. A task can be paused and resumed from the same point — ax suspend and ax resume — and via ax ssh you can go inside the sandbox and see what the agent is doing.

The CLI talks to the control layer over gRPC and follows the active Kubernetes context: when you switch clusters, AX finds and tunnels to the right control plane on its own. Deployment requires a Kubernetes cluster, the ko utility, an accessible container registry and the Agent Substrate Control API — everything installs into the ax-system namespace.

The stated goal is billions of tasks in a single cluster, hence the bet on high throughput. The code is open under the Apache 2.0 license, but the project is still in alpha: the developers warn of possible major breaking changes before a stable release.

Image: Grendelkhan CC BY-SA 3.0 · License

Translation editor: Салтанов А.