News · 2026-09-21

Google-originated AX opens an execution control plane for agents

Google-originated AX is now an open-source control plane for running isolated, stateful AI-agent workloads on Kubernetes. Its important contribution is not a clever agent prompt or model: it is the operational machinery that decides where an agent runs, what it can reach, how its workspace is prepared and how it pauses or resumes.

Key facts

• The project is published as google/ax under Apache-2.0.
• Google Cloud announced the related Agent Executor runtime on 20 May 2026.
• AX defines Task, Workspace, Gateway and Model resources for agent execution.
• Its README claims “billions” of tasks per cluster, an official claim not independently benchmarked.

AX is best understood as the ground-control system around an agent. A Task describes an isolated execution unit, including image, command, resource limits and workspace references. A Workspace can bootstrap Git repositories, MCP servers and skill registries. A Gateway defines listeners and outbound host/port allowlists. A Model centralizes provider and secret configuration; it is not itself a model. This separation matters because a capable coding or research agent becomes much safer and more reproducible when its tools, filesystem and network boundary are explicit rather than improvised.

The concepts document makes clear that a task can be a whole job or a node in a tree of child tasks. The design document says AX avoids placing huge numbers of short-lived tasks in Kubernetes custom resources, which can strain etcd. Instead it uses Redis for state and Redis Streams between an API server and scaled controllers. Those controllers provision execution spaces and actors through Agent Substrate; a task runner materializes the workspace and launches the configured command.

The concrete analogy is a container terminal. The cargo—an agent’s reasoning and tool decisions—comes from somewhere else. AX supplies the cranes, manifests, gates, isolation lanes and suspension rules that keep many containers moving without allowing each one to reach every part of the port. That is why its egress policy and workspace materialization are more consequential than its model configuration field.

Google’s own launch post says the design draws on work with Google DeepMind and was developed with the Google Kubernetes Engine team. That supports Google Cloud involvement. It does not make AX an officially supported DeepMind product or prove the company will build its agent strategy around it. The separate Agent Substrate repository explicitly says it is not an officially supported Google product, a distinction often lost in social discussion.

The project’s scale claim needs the same care. AX says it can run billions of autonomous workloads per cluster; Google’s Agent Substrate material discusses hundreds of millions of registered agents and millions of sub-second tool calls. A published substrate demo shows roughly 250 stateful actors multiplexed across eight physical pods. These support the design’s ambition and state-preservation mechanism, but they are not an independent proof of billion-task production use.

The strongest counterargument is practical: Kubernetes is operationally expensive for a small team whose agents could run in a VM, worktree or managed sandbox. The strongest rebuttal is that isolation, scoped networking, secrets and lifecycle become first-class constraints at fleet scale. AX is pre-stable and should be evaluated as infrastructure, not adopted because of an unverified scale slogan. Its emergence is still a marker of where agent engineering is heading: the harness is becoming a distributed systems problem.

Primary source, verified: read the paper →

Ground Truth. (2026, September 21). Google-originated AX opens an execution control plane for agents. Ground Truth. https://groundtruth.day/news/google-ax-agent-executor-open-source.html

@misc{groundtruth:google-ax-agent-executor-open-source,
  title  = {Google-originated AX opens an execution control plane for agents},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {sep},
  url    = {https://groundtruth.day/news/google-ax-agent-executor-open-source.html}
}

Topics: agents · infrastructure · google · open-source · kubernetes