Google AX

Overview

AX (Agent Executor) is Google’s open-source distributed runtime for orchestrating autonomous AI agent workloads. Rather than providing agent logic or prompt engineering tools, AX sits at the infrastructure layer — it handles sandboxing, workspace pre-wiring, network policy enforcement, and lifecycle management so that platform teams can run untrusted agent code safely and at scale. The project was announced at Google I/O on September 18, 2026, and immediately hit the top of Hacker News with over 600 points.

Key Capabilities

AX uses a declarative model inspired by Kubernetes. You define a task in a YAML manifest specifying the agent image, its workspace (Git repos, MCP servers, skill packages), gateway rules (which external hosts the agent can reach), and the LLM provider and model. The runtime then provisions a sandboxed container, wires up the workspace, enforces network fencing, and manages execution.

The standout feature is checkpoint-based lifecycle management. Agents can be suspended when idle and resumed from checkpoint, eliminating compute costs during wait states and making multi-hour autonomous runs survivable across restarts and failures. Trajectory branching lets operators rewind to any checkpoint and test alternative execution paths — useful for debugging agent decisions or evaluating strategy variants.

The control plane uses a single-writer architecture with an append-only event log. A central controller mediates all interactions for auditing and policy enforcement. The CLI provides interactive debugging via ax ssh, letting operators shell into running agent containers.

Use Cases

AX targets platform engineering teams running fleets of autonomous agents — CI/CD pipelines with agent-driven code review, security scanning, infrastructure management, and research automation. Its Kubernetes-native architecture makes it a natural fit for organizations already running K8s clusters. The checkpoint and suspend features are particularly valuable for long-running agent workflows where idle compute costs add up.

Considerations

AX requires a Kubernetes cluster, container registry, and Google’s Agent Substrate runtime. This is not a lightweight local tool — it is enterprise infrastructure. The project is in alpha with breaking changes expected, so production adoption should be cautious. It is model-agnostic but examples default to Gemini. The Go codebase is well-structured but documentation is early-stage.

Who It’s For

AX is built for platform engineers, DevOps teams, and infrastructure architects who need to run many autonomous agents safely in a cluster. It appeals to enterprises with existing Kubernetes infrastructure who want production-grade agent orchestration with security guarantees. Individual developers building single-agent applications should look at higher-level frameworks instead.