Google Releases AX v0.3.0, an Agent Task Orchestrator for Kubernetes

Google has published AX v0.3.0 in the google/ax repository. The open project provides a Kubernetes-native layer for orchestrating, isolating, and managing environments for agent tasks, but its API may still change substantially.

Google AX v0.3.0 is a new version of an open orchestrator for running agent tasks in Kubernetes. The release is available in the public google/ax repository, which describes the project as “Google’s open agentic orchestrator” and distributes it under the Apache-2.0 license.

Version v0.3.0 was released on September 20, 2026. It is a recent update to a project whose first release, v0.1.0, was published in May. AX does not focus on creating the AI agent’s underlying logic, but on the operational layer: workspaces, execution-code isolation, access to tools and models, task state, and network-communication restrictions.

Google AX v0.3.0 Is Built on Declarative Manifests

AX uses declarative manifests to describe and manage agent workloads. The documentation lists four main primitives: Task, Workspace, Gateway, and Model.

  • Task represents an agent task being executed.
  • Workspace serves as a working environment for tasks, including isolated sandboxes.
  • Gateway covers access to external services and tools.
  • Model is used to configure the models in use.

According to the documentation, the project supports preparing repositories and MCP servers in the workspace. It also includes rules for outgoing network communication. These features are relevant for agents that work with code, external tools, or untrusted inputs and need a separate runtime environment.

AX requires Kubernetes and Agent Substrate. Teams that want to try it must therefore plan for deployment in a Kubernetes environment; it is not a standalone local library for building agents.

An Early Project with Unstable Interfaces

The most important limitation is the project’s status. The documentation explicitly warns that substantial breaking changes are expected before a stable release. Interfaces and integration methods may therefore change between versions in ways that require modifications to existing deployments.

It has also not been confirmed that AX is ready for production deployment. The public repository in the google organization confirms publication under that account, but the available materials do not specify the product owner, support level, or Google’s long-term plans.

The project website makes claims about scaling to very large numbers of agent tasks or sessions, fast recovery, and denser multiplexing. However, no independent benchmarks have been provided for these parameters, so they cannot yet be compared with existing Kubernetes or agent runtimes.

What to Watch Next

For development and platform teams, AX may be interesting as an open Kubernetes-native layer for managing agent workspaces. Given its current status, however, it makes sense to focus primarily on the release pace, API stabilization, and any removal of the breaking-changes warning.

Independent tests of performance, isolation, and operating costs will also be important. Another open question is whether Google will publish a separate official announcement, roadmap, or project support terms.

Sources

  • GitHub — google/ax README – Confirms the public repository, license, AX’s declared features, its dependence on Kubernetes and Agent Substrate, and the warning about unstable interfaces.
  • GitHub — google/ax Releases – Confirms the release of version v0.3.0 on September 20, 2026, and earlier project releases, including the initial v0.1.0.
  • AX – Describes the project’s intended use, its primitives, and its own claims about task scaling and recovery.

Verified and updated: 09/21/2026 06:23

Tags
Sharing

Súvisiace