Google shipped the Kubernetes of agents and called it AX
Google just open-sourced AX, short for Agent Executor, and the pitch is not subtle: run billions of autonomous agent workloads per cluster. It went up at github.com/google/ax under Apache 2.0, hit the top AI slot on Hacker News with over 600 points, and the landing page is agentexecutor.io.
The design is deliberately kubectl-shaped. Four primitives and you are done: Task is an isolated sandbox that runs the work, Workspace wires up the environment, Gateway fences the network, Model handles configuration. You apply, get, describe, watch, delete, plus a few agent-specific verbs. Anyone who has run a cluster will be productive in about ten minutes, which is obviously the point.
Underneath it sits Agent Substrate, a compute runtime Google built for actor density rather than for containers. That is where the numbers come from: sub-second resumption, dozens of tasks multiplexed onto shared worker resources, and the feature everyone in the thread actually latched onto, trajectory branching from kernel snapshots. Fork a running agent's entire state and explore two futures from the same point. That is not a scheduling feature, that is a debugging and search feature, and nobody else is shipping it at this layer.
The v0.3.0 release that landed with it tells you where the real engineering went. AX split into three services, an API frontend, a reconciler, and a sandboxed task runner, and moved task state out of Kubernetes custom resources into Redis Streams. Translation: they hit the wall where etcd stops being a reasonable place to keep millions of mutable agent states, and they routed around it. That is a very specific kind of scar tissue and it only comes from actually running the thing.
Here is why this matters more than the average Google GitHub drop. The repo lives under the official google org with a real license, which commenters correctly flagged as the reason to take it seriously. And the framing is a bet: Google thinks the scarce resource in agents is not the model, it is the control plane. Dynamic scheduling, resumption, auto recovery, auditing, all treated as first-class rather than as something you bolt on after your agent fleet catches fire. Every team currently gluing agent orchestration together out of a queue, a retry loop and hope should read this before writing another line of that glue. https://github.com/google/ax
← Back to all articles
The design is deliberately kubectl-shaped. Four primitives and you are done: Task is an isolated sandbox that runs the work, Workspace wires up the environment, Gateway fences the network, Model handles configuration. You apply, get, describe, watch, delete, plus a few agent-specific verbs. Anyone who has run a cluster will be productive in about ten minutes, which is obviously the point.
Underneath it sits Agent Substrate, a compute runtime Google built for actor density rather than for containers. That is where the numbers come from: sub-second resumption, dozens of tasks multiplexed onto shared worker resources, and the feature everyone in the thread actually latched onto, trajectory branching from kernel snapshots. Fork a running agent's entire state and explore two futures from the same point. That is not a scheduling feature, that is a debugging and search feature, and nobody else is shipping it at this layer.
The v0.3.0 release that landed with it tells you where the real engineering went. AX split into three services, an API frontend, a reconciler, and a sandboxed task runner, and moved task state out of Kubernetes custom resources into Redis Streams. Translation: they hit the wall where etcd stops being a reasonable place to keep millions of mutable agent states, and they routed around it. That is a very specific kind of scar tissue and it only comes from actually running the thing.
Here is why this matters more than the average Google GitHub drop. The repo lives under the official google org with a real license, which commenters correctly flagged as the reason to take it seriously. And the framing is a bet: Google thinks the scarce resource in agents is not the model, it is the control plane. Dynamic scheduling, resumption, auto recovery, auditing, all treated as first-class rather than as something you bolt on after your agent fleet catches fire. Every team currently gluing agent orchestration together out of a queue, a retry loop and hope should read this before writing another line of that glue. https://github.com/google/ax
Comments