Google Open-Sources AX: A Distributed Agent Runtime Built for Reliability and Resumption

AIDeveloper44 Team

What Is AX?

Google has quietly released AX (Agent eXecutor), an open-source distributed agent runtime. The project is in active early development — Google notes it will introduce major breaking changes before a stable release — but the architecture and capabilities are already substantial.

AX provides the runtime layer that's been missing from most agentic AI stacks: a way to coordinate multi-step agentic loops reliably, with full auditability, automatic recovery from failures, and the ability to resume interrupted executions from exactly where they left off.

Core Architecture

AX is built around a single-writer controller that coordinates all execution. The key components are:

• Controller — manages the agentic loop, event log, and registry; single-writer architecture ensures consistent state
• Remote Agents — isolated actors implementing the AgentService RPC, spawned and coordinated by the controller
• Tools — MCP servers providing skill execution
• Environment — isolated actor with built-in tools and skills
• Router — handles resumable streams between client and controller

Resumption as a First-Class Feature

What makes AX distinctive is its treatment of execution resumption. In most agent frameworks, a network interruption or node failure means starting over. AX maintains a durable event log and supports:

• Automatic recovery — agents resume from their last durable state after failures
• Client catch-up — a disconnected client can replay missed events by passing its last sequence number
• Conversation branching — ax fork lets you branch from any checkpoint, enabling experimentation without losing prior work
• Compute-layer actor resumption — on compatible platforms (Kubernetes), individual actors can be resumed at the infrastructure level

Getting Started

AX is written in Go and installs via:

go install github.com/google/ax/cmd/ax@latest

It's model-agnostic and harness-agnostic by design. For production use, Google recommends deployment on Kubernetes via Agent Substrate, which provides higher-density scheduling for agentic workloads. Custom agents can be registered as remote gRPC services or sandbox agents.

Google is building AX in public specifically to validate design decisions before cutting a stable release — feedback is actively solicited at ax-dev@google.com. Explore the repository at github.com/google/ax.

Enjoyed this?

Get more posts like this delivered to your inbox.

Subscribe