Idle waiting is where agent infrastructure bleeds money, and Google’s answer is to freeze sessions mid-task. The company open-sourced AX this week, an Apache 2.0 runtime living at agentexecutor.io and in a GitHub repository under google/ax, running on an execution layer called Agent Substrate.
The complaint it addresses is structural. A coding agent alternates bursts of reasoning and tool use with long stretches waiting on a model or an API, a poor fit for container orchestration built around short request-response cycles. Keeping sandboxes warm wastes compute, while letting them go cold adds latency to every loop. AX gives each session an isolated actor sandbox with hard CPU and memory limits, checkpoints state when the agent goes quiet, and resumes in under a second. Dozens of tasks then multiplex onto shared host workers.
Four primitives make up the control plane, all under the ax.io/v1alpha1 group. Task sets the lifecycle and resource ceilings. Workspace assembles the environment first, mounting Git repositories, configuring Model Context Protocol servers, installing skill bundles, or handing a natural-language goal to an initialization agent. Gateway enforces outbound hostname allowlists. Model registers inference providers. Systems researchers across Google and Google DeepMind fed the design.
