Google-origin AX platform targets large-scale AI-agent orchestration
AX is described as a declarative control plane for running stateful AI-agent workloads in isolated sandboxes, with suspend-and-resume execution and infrastructure managed through configuration.
AX has been presented as an open, declarative platform for running AI-agent tasks in isolated, resource-limited sandboxes. The product description says it provisions workspaces, network policies, model configurations and secrets before an agent starts work.
The platform is designed around workloads that accumulate state, make calls to model APIs and external tools, and may alternate between periods of intensive computation and idle waiting. AX says tasks can be suspended while awaiting model, tool or human responses, then resumed rapidly to share worker capacity.
AX describes each task as a lightweight actor running on Agent Substrate, a compute runtime for stateful actor lifecycles. Its product materials claim the architecture is intended to support very large numbers of concurrent sessions, although those scale and performance claims have not been independently verified.
The platform supports interactive coding agents, long-running agent servers, Jupyter notebooks, browser testing and custom tool runtimes, according to the source. It also targets trajectory collection, reinforcement-learning loops and agent evaluation through reproducible sandboxes.
AX says it emerged from agent-runtime research and large-scale infrastructure work at Google, including Google DeepMind, and describes itself as an open control plane for agent execution. The supplied material does not provide licensing details, independent users, benchmarks or evidence of production deployment, and does not independently confirm Google’s involvement.
