Google open-sources AX, an AI agent runtime for building and managing agents without code

Google says AX started as research, but the open-source runtime is now pitched as production infrastructure for teams that need to create and manage AI agents at scale. AX runs each task in a lightweight actor, supports billions of concurrent agent sessions per cluster, and sandboxes every task before it starts. It also sets up the workspace, fences network access, and can checkpoint idle agents while they wait for model responses, external tool calls, or human input.

The system is designed to resume those paused agents in under a second. Google also says AX uses dense multiplexing, which turns idle waiting time into spare compute capacity instead of leaving it unused.

Another notable part of the platform is that a workspace can be described in plain English and prepared automatically before the task begins. AX supports interactive coding agents, long-running agent servers, Jupyter notebooks, headless browser testing, and custom tool runtimes.

Google says AX runs on top of Agent Substrate. On the project site, the company also describes agents as a different kind of workload: not microservices and not batch jobs.

This kind of runtime raises a simple question: if agents become a standard workload, how quickly will developers adopt systems built for them from the start?