Skip to content

Back to the catalog

github.comrepo

google/ax — declarative agentic orchestration runtime

AX is Google's declarative orchestrator for running large numbers of autonomous agent workloads on Kubernetes, built on top of "Agent Substrate" for sandboxed execution. Agents are declared via YAML manifests (Task, Workspace, Model) similar to kubectl, and applied with an `ax` CLI that supports apply/get/watch/ssh/suspend/resume, treating agent runs as a distinct workload type distinct from stateless services or batch jobs.

TL;DR Google's Kubernetes-style declarative orchestrator for running agent workloads at scale, built on Agent Substrate sandboxing.
Open original ↗
Document

How to use it

Two installs, in order. Agent Substrate must already be running — it lands in the ate-system namespace and exposes its Control API at api.ate-system.svc.cluster.local:443, which is where AX looks for it. Verify with kubectl get svc api -n ate-system before continuing. Then install the CLI and deploy the control plane:

go install github.com/google/ax/cmd/ax@latest
make deploy AX_IMAGE_REPO=<your-registry>

make deploy stands up Redis, then builds and pushes the control-plane images with ko (so you need ko and a cluster-reachable registry), landing everything in the ax-system namespace. Prerequisites beyond that are Go and kubectl.

The ax CLI talks to the control plane over gRPC and follows your active Kubernetes context — switching contexts (kubectx prod-cluster) makes ax resolve and tunnel to that cluster's control plane in the background; ax --context=dev-cluster get tasks targets one without switching. Tunnel state lives in ~/.ax/tunnels.

Beyond the verbs the summary lists, note the semantics: ax suspend checkpoints actor state and pauses; ax resume picks up from that checkpoint. ax ssh requires spec.debug: true on the task. Run ./demo.sh for a full lifecycle (custom workspace → readiness → ssh commands → suspend).

The three manifest kinds carry more than the summary implies: a Workspace wires up git repos and MCP servers (ax get workspaces shows GIT-REPOS and MCP-SERVERS columns); a Model names a provider and model (the example default is provider: google, model: gemini-3.8-flash).

Limits

The README carries an explicit warning: core concepts, protocols, and specifications are still being refined, and major breaking changes are expected before a stable release — the API is ax.io/v1alpha1. This is not production-stable.

The "billions of autonomous agent workloads per cluster" throughput claim is a design goal stated on the page, not a benchmarked result — no numbers back it.

Hard dependency: AX does nothing without Agent Substrate installed and its Control API reachable first. License is Apache 2.0.

Whether this is a runnable system or an early Google research/infra release is worth checking against the alpha warning before you build on the manifest schema — it will move.

Organisations named here

  • Google

Technologies named here

  • AX
  • Gemini 3.8 Flash
  • Agent Substrate
  • Kubernetes
  • kubectl
  • ko
  • Redis
  • kubectx
  • Docker
  • gRPC
  • Go

Techniques named here

  • Task suspend/resume
  • Declarative manifest-based orchestration
  • Debug/SSH sandbox access

Related in agents