Google open-sources AX to orchestrate millions of AI agents like K8s

AX (Agent Executor) is Google's open-source agentic orchestration runtime designed to declare, isolate, and run millions of autonomous AI agent workloads across a Kubernetes cluster. If you already know your way around kubectl, AX is built to feel immediately familiar.

Autonomous agents are fundamentally different from traditional stateless microservices or batch jobs. They accumulate conversational state, call third-party APIs and tool servers, require hard sandboxing, and will happily burn thousands of dollars looping forever if unattended. AX tackles this by turning agent operations into declarative infrastructure.

What it does

AX boils agent orchestration down to three core primitives: Task, Workspace, and Model.

A Task packages untrusted agent code into an isolated sandbox configured with strict CPU and memory limits. A Workspace pre-wires required Git repositories, MCP tool servers, and skill packages so agents start warm instead of downloading dependencies every time. A Model defines cluster-level LLM connections, letting the control plane handle credentials through native Kubernetes secrets.

The real magic is in day-to-day operations. The ax CLI talks over gRPC and mirrors standard cloud-native workflows. You apply tasks via YAML, inspect them with ax get and ax describe, and seamlessly track lifecycle phases.

Two verbs stand out for debugging and cost control. With ax suspend and ax resume, you can checkpoint an idle or looping agent to disk and wake it up later without losing progress. And when an agent behaves strangely, ax ssh lets developers drop straight into the sandboxed container to inspect files and run test commands in real time.

How it runs

AX is licensed under Apache 2.0 and hosted on GitHub under Google's repository. Setting it up requires a fairly robust infrastructure baseline.

AX schedules tasks as actors on top of Agent Substrate, which must already be deployed and serving its Control API inside your Kubernetes cluster's ate-system namespace. On your management machine, you will need Go, kubectl, and Podman or Docker with authenticated registry access.

After installing the CLI with go install, you build and push the server and runner container images, then deploy the control plane and PostgreSQL database into the ax-system namespace. Once deployed, the ax CLI automatically detects your active kube context and tunnels traffic to the cluster.

Google notes that AX is under heavy active development. Core concepts, communication protocols, and specs will likely undergo major breaking changes before reaching a stable release.

Why it matters

Most agent frameworks focus on prompt chaining and memory abstractions, largely ignoring what happens when thousands of autonomous tasks run concurrently. AX treats agents like actual Linux workloads that need sandboxing, process supervision, and hard cost boundaries.

Platform engineers and infra teams building enterprise agent fleets should inspect the architecture and primitives today, as declarative orchestration is clearly where the field is heading. That said, production adopters should hold off for now: running pre-alpha software on top of custom cluster substrates is strictly for early builders willing to stomach breaking updates.