google/ax

What it solves

AX is designed to handle the unique requirements of autonomous agents, which differ from standard microservices or batch jobs. It provides a way to run agentic workloads at scale in a cluster, ensuring they are isolated in sandboxes to prevent untrusted code from causing harm, and managing their state so they can be paused and resumed without losing progress.

How it works

AX operates as a declarative orchestrator built on top of Agent Substrate. It uses a Kubernetes-like approach where users define their agent's environment and requirements using YAML manifests.

It utilizes three primary primitives:

• Task: Defines an isolated sandbox with CPU and memory limits to run agent code.
• Workspace: Pre-configures the environment with Git repositories, MCP servers, and skill packages so agents start "warm."
• Model: Configures the LLM and credentials used by the platform.

Users manage these tasks via a CLI (ax) that supports commands like apply to deploy, watch to monitor, and ssh to inspect the running sandbox.

Who it’s for

Developers and platform engineers who need to deploy and manage billions of autonomous agent workloads across a Kubernetes cluster with strict isolation and state management.

Highlights

• Declarative Orchestration: Uses YAML manifests to define tasks, workspaces, and models.
• ció-Kubernetes Experience: CLI tools that mirror kubectl for easy adoption by DevOps engineers.
• State Persistence: Ability to suspend and resume agents to save costs and resources.
• Sandboxed Execution: High-throughput execution of untrusted code via Agent Substrate.
• Direct Inspection: ax ssh allows developers to shell into a running agent's sandbox for debugging.