Google has published ax, an open source orchestrator designed to run large numbers of autonomous agent workloads inside a Kubernetes cluster, according to the project’s README on GitHub. Each task runs in an isolated sandbox with its own workspace, and a network gateway restricts outbound traffic to an explicit allowlist, addressing agents that call external model APIs in a loop and can burn money unsupervised.
The system builds on Google’s Agent Substrate for sandboxed execution and exposes four primitives, Task, Workspace, Gateway, and Model, defined through YAML manifests. A single command line tool, deliberately shaped like kubectl, applies those manifests. Engineers can pause a running agent and resume it later, or shell directly into a live sandbox to inspect what it is doing.
Google’s README flags ax as unstable, warning that breaking changes to its core concepts are likely before a stable release. The repository is licensed under Apache 2.0.
Per the project’s README on GitHub (Google), accessed September 21, 2026.
