Google AX is an open-source agent harness and execution orchestrator developed by Google. It provides runtime infrastructure and execution sandboxing designed to bridge terminal-based tool environments with Google's Antigravity agent architecture and Gemini/Vertex AI model backends. Day-to-day, AX operates as a command-line harness (`cmd/ax`) interfacing with a local sidecar service. The system manages conversational state across interactions, maintaining persistent execution directories (`AX_HARNESS_WORKDIR`) and routing tool executions such as shell commands, code generation, and repository edits. It communicates with language models primarily through Google Cloud Vertex AI and the Google GenAI SDK ecosystem. What distinguishes AX from conventional standalone CLI agents is its decoupled harness architecture. Instead of tightly coupling model calls with the user terminal, AX implements a client-sidecar model that enforces workspace boundaries, isolates execution environments (including support for sandboxed/gVisor environments), and cleanly serializes conversation turns to durable storage for session resumption. Google AX is open-source software distributed under an Apache 2.0 license. Model inference costs depend on your underlying Google Cloud Vertex AI or Google AI Studio billing configuration.
Tags: agent harness, AI, ai agent, cli, developer tools, OpenSource