Google introduces AX, an open agentic orchestrator for sandboxed, stateful AI tasks that can run at massive scale

Google’s AX lets developers declare agentic tasks in YAML, then runs them in isolated sandboxes with configured workspaces, network controls, and suspend/resume support. The system is designed for agents that accumulate state, use tools and model APIs, and need guardrails to avoid runaway cost or unsafe execution.

AX reflects a shift toward treating AI agents as a distinct infrastructure workload rather than as microservices or batch jobs. If adopted, agent orchestration platforms like this could become a core layer for safely scaling autonomous coding, research, and operations agents.