Back to News

### TL;DR

AX (Agent Executor) is Google's open-source distributed agent runtime for production AI agents. It coordinates agentic loops across distributed isolated actors (skills, tools, agents), manages durable execution state with automatic recovery and resumption, and provides a single-writer architecture for consistent state management — designed to scale from single machines to Kubernetes clusters.

Key Insights & Metrics

Key Features

• Distributed agent runtime — controller, skills, tools, and agents execute in isolation as separate actors; event log ensures durable execution state with automatic recovery and resumption even in complex distributed setups; supports branching from any checkpoint via ax fork
• Built-in resumption and reliability — single-writer architecture ensures consistent state; clients can reconnect mid-execution and replay missed events; native Kubernetes support via Agent Substrate with higher-density agentic workload scheduling
• Model and harness agnostic — supports custom remote agents via gRPC AgentService, MCP tool servers, and any LLM backend; complete auditing and policy control of all user and agentic calls through a central controller; CLI for local or server execution

Discussion