Back to wire
Development·Article·Confirmed

Google's AX 0.3.0 packages agent fleets into a Kubernetes-style control plane

Google has released AX 0.3.0, an Apache-2.0 control plane that models agent workloads as Task, Workspace, Gateway and Model resources over Agent Substrate. It provides sandboxing, repository and MCP pre-wiring, outbound-network allowlists and suspend/resume, while the project remains explicitly pre-stable and its claimed billion-task scale has not been independently reproduced.

Published 20 Sept 2026, 05:33 · Updated 21 Sept 2026, 12:53

AX turns agent jobs into declarative resources

Google's AX project published version 0.3.0 on 20 September. The current design is a Kubernetes-shaped control plane for long-running agent workloads, with four first-class resources: Task for execution, Workspace for repositories and tools, Gateway for network policy, and Model for the language model configuration used by the platform.

The command-line interface deliberately follows kubectl conventions. Operators apply ax.io/v1alpha1 manifests, inspect and watch tasks, open an interactive shell into debug-enabled sandboxes, and can suspend a running agent before resuming it from its saved state. The project is written in Go and released under Apache-2.0.

Isolation and network policy sit inside the orchestration model

Tasks run on Agent Substrate, the sandbox layer AX uses to isolate untrusted agent code and apply CPU and memory limits. A Workspace can pre-wire Git repositories, MCP servers and skill packages before an agent starts, while a Gateway can restrict outbound traffic to an explicit host allowlist instead of giving every task unrestricted internet access.

Model credentials are supplied through Kubernetes secrets rather than being embedded in a task manifest. Taken together, those primitives target a practical difference between ordinary batch jobs and coding or research agents: agents accumulate state, call external tools, need scoped credentials and may continue consuming compute if a loop is left unconstrained.

The source is inspectable, but deployment still needs a real cluster

AX can be installed from source with Go, and its documentation includes manifests, a lifecycle demo and design references for the control plane. Running the system requires a Kubernetes cluster, ko, a container registry and a reachable Agent Substrate control API. The 0.3.0 GitHub release contains source archives rather than prebuilt release assets, so reproducing the documented path still means building and deploying the stack rather than downloading a self-contained binary distribution.

The repository remains active: v0.3.0 is five commits ahead of v0.2.3, including runtime and dependency work, and the repository was pushed again on 20 September. GitHub showed 4,346 stars and 190 forks when checked on 21 September; those figures are dated adoption signals, not evidence that the system has been proven at production scale.

Pre-stable status limits the scale claim

Google's README explicitly warns that AX is still refining its core concepts, protocols and specifications and is likely to make major breaking changes before a stable release. The API resources are v1alpha1, which makes compatibility guarantees and upgrade behaviour important constraints for anyone considering the project as an infrastructure dependency today.

The project also says AX is built to run billions of tasks per cluster. No public benchmark, reproducible load test or independent deployment result currently establishes that scale. The confirmed development is narrower: Google maintains an Apache-2.0 agent orchestration stack with inspectable scheduling, sandbox, workspace and network-policy mechanisms, and version 0.3.0 is now the latest tagged release.

Source trail

3 sources · 3 primary