AX is Google's open-source agent orchestration runtime. Google officially describes it as a high-throughput, declarative orchestrator for running billions of autonomous agent workloads across a cluster. It is built on top of Agent Substrate, leveraging its sandboxed execution capabilities, and is designed to run billions of tasks in a single cluster. If you have used Kubernetes, the AX experience will feel quite similar.
In terms of purpose, AX targets a new type of workload: agents. They are neither stateless microservices nor batch jobs that run to completion; instead, they accumulate state, require strict isolation, call model APIs and tool servers, and may fall into loops that burn money if left unmonitored. AX provides three small primitives to handle these requirements declaratively. Everything is expressed as ax.io/v1alpha1 manifests and applied with a single command. The workflow is roughly: declare a task, apply it, watch it start, and view it from the agent's perspective.
Key capabilities include: declaratively defining agent tasks (including workspace and model specifications); scheduling tasks as sandboxed actors onto Agent Substrate; supporting CPU/memory limits for tasks; providing suspend and resume; supporting debugging via ax ssh; a command-line interface with a kubectl-like shape, including apply, get, describe, watch, delete, and several agent-specific verbs, and compatible with kubectx. ax communicates with the control plane via gRPC and follows the currently active Kubernetes context, resolving and tunneling to that cluster's control plane in the background.
In terms of usage, several prerequisites must be met before deployment: a Kubernetes cluster with Agent Substrate installed; ko installed (via brew install ko) and a container image registry that the cluster can pull from. First install Agent Substrate according to the Substrate README instructions. It will be deployed in the ate-system namespace and expose the Control API at api.ate-system.svc.cluster.local:443, which is where AX looks by default. After installing the CLI, the ax binary will be placed in $(go env GOPATH)/bin, so make sure that directory is in PATH. Then deploy the control plane: this step first deploys Redis, then uses ko to build and deploy the control plane image, with everything landing in the ax-system namespace. Finally, you can run your first task; if you want to see the full lifecycle end to end, run ./demo.sh, which applies a custom workspace, waits for readiness, executes commands via ax ssh, and suspends the task.
The applicable scenario is running autonomous agent workloads at scale in a Kubernetes cluster. Limitations include: the project is still under heavy development, and core concepts, protocols, and specifications are being actively refined, so major breaking changes are likely before a stable release; Agent Substrate must already be running in the cluster before running AX. The project is licensed under the Apache License 2.0, with details in the repository's LICENSE file.
