01
Overview
Google Agent Executor (AX) is Google's open-source runtime for orchestrating autonomous agent workloads. Google originally introduced Agent Executor in May 2026 as a distributed runtime for durable, resumable agent execution. The September 20 v0.3.0 release is a more useful SEO event because the project now presents a materially redesigned, Kubernetes-shaped orchestration model.
02
What changed in AX v0.3.0?
The current official repository describes AX as a high-throughput declarative orchestrator. Instead of treating an agent as a single opaque process, AX exposes four core resources through ax.io/v1alpha1 manifests: Task, Workspace, Gateway and Model.
A Task is the isolated execution unit. A Workspace defines what the task needs, including Git repositories, MCP servers and skills. A Gateway controls network access, including explicit outbound hosts and ports. A Model centralizes model configuration and credentials. This makes AX closer to an infrastructure control plane for agents than to a prompt or agent-framework library.
03
Kubernetes-style workflow
The AX CLI deliberately resembles kubectl. Developers apply YAML with ax apply, inspect resources with get and describe, stream state with watch, and remove resources with delete. Agent-specific commands include ax suspend, ax resume and ax ssh.
That distinction matters: AX does not replace the agent's reasoning framework. It provides the runtime and lifecycle layer around the code, model calls, tools and environment.
04
Task, Workspace, Gateway and Model
Task runs agent code in an isolated sandbox with CPU and memory limits.
Workspace prepares the environment before the task starts. The current project can pre-wire Git repositories, MCP servers and skills, and also supports goal-based generative workspace setup.
Gateway is the network-policy boundary. It can restrict outbound traffic to an allowlist, which is especially relevant when autonomous code can call external services.
Model centralizes the model/provider configuration and credentials used by the platform.
05
Agent Substrate relationship
AX runs on top of Agent Substrate, a separate open-source compute runtime designed for high-density, stateful agent workloads. Google's project materials describe AX as the declarative control plane and Agent Substrate as the execution substrate underneath it.
The project claims it is designed to scale to billions of tasks per cluster and describes sub-second resumption. Those are project/design claims, not independently verified production benchmarks, so they should not be presented as guaranteed capacity for every deployment.
06
Installation and requirements
The current quick start installs the CLI with Go. Deploying the control plane requires a Kubernetes cluster, ko, a container registry the cluster can pull from, and a reachable Agent Substrate Control API. The deployment installs Redis and AX control-plane components in the ax-system namespace.
Because the project is explicitly pre-stable, exact commands and dependencies should be rechecked against the repository immediately before publication.
07
Why AX is interesting for AI agents
Long-running agents create infrastructure problems ordinary request/response applications do not: they can wait for model calls or humans, accumulate state, execute untrusted code, need controlled network access and consume resources while idle. AX is Google's attempt to make those concerns first-class runtime primitives rather than application-specific glue.
Potential uses include coding agents, research agents, evaluation farms, browser-testing workers, notebook workloads and custom tool runtimes.
08
Security boundaries
AX exposes several useful controls: isolated tasks, CPU/memory limits, Gateway egress policies and centralized model credentials. ax ssh also provides debugging access when a task is explicitly configured for debug mode.
These features are infrastructure controls, not proof that arbitrary agent code is safe. Teams still need to reason about credentials, tool permissions, supply-chain risk, model behavior and the security of the underlying Kubernetes and Agent Substrate deployment.
09
Is Google AX open source?
Yes. The google/ax repository is public and licensed under Apache 2.0.
10
Is AX stable?
No. The official repository warns that core concepts, protocols and specifications are still being refined and that major breaking changes are likely before a stable release.
11
Is AX a hosted Google Cloud service?
The reviewed sources describe AX as open-source software that developers deploy on their own infrastructure. Do not infer a standalone hosted-product price from that. Infrastructure cost depends on the chosen cluster, compute, model providers and surrounding services.
12
AX vs Kubernetes
AX uses Kubernetes-like concepts and runs in a Kubernetes ecosystem, but it adds agent-specific abstractions for sandboxed tasks, prepared workspaces, network gateways, model configuration and suspend/resume. A dedicated comparison page is not yet necessary because this distinction can be answered well inside the main overview.
13
FAQ
What does AX stand for?
AX is the short name for Agent Executor.
Can AX run MCP servers?
The current Workspace abstraction explicitly includes MCP servers among the resources that can be prepared for a task.
Can AX pause an agent?
The CLI documents ax suspend and ax resume for checkpointing and resuming tasks.
Does AX replace LangGraph or other agent frameworks?
Not necessarily. AX is primarily an execution/orchestration runtime. Agent logic can live in the harness or application running inside the task.
Where is AX available?
The source code and documentation are publicly available through the official google/ax repository and agentexecutor.io.
Sources
Primary and supporting sources
Use the linked documentation to check current pricing, access and product behaviour. The dates above describe this article; they do not guarantee that every vendor claim remains current.