Google Releases Declarative Agent Orchestrator 'AX'

Google has released AX, an open-source orchestration runtime for declaratively managing agent workloads. It offers a user experience similar to Kubernetes, handling sandboxing, network isolation, and resource limits when agent tasks are declared.

4 Core Primitives

AX provides the following four primitives for agent execution:

• Task: Executes untrusted agent code in isolated sandboxes with CPU/memory limits
• Workspace: Pre-connects Git repositories, MCP servers, and skill packages to provide a ready state at agent startup
• Gateway: Restricts outbound traffic to an explicit host allowlist
• Model: Retrieves credentials from Kubernetes Secrets to configure the LLM used by the platform

CLI and Deployment

All configurations are expressed as YAML manifests using the ax.io/v1alpha1 API version and applied via the ax apply command. The CLI supports a form similar to kubectl (apply, get, describe, watch, delete), supports suspending and resuming idle agents with ax suspend/resume, and supports shell access to running agents with ax ssh. Control plane deployment requires a Kubernetes cluster, ko, a container registry, and the Agent Substrate Control API.

Caveats and Status

Core concepts and protocols are currently being actively refined, and major breaking changes may occur before a stable release. The project follows the Apache License 2.0 and is under public development on GitHub.