Google-Born AX Aims to Scale Agentic Workloads

As agentexecutor.io first reported, Google is introducing AX, an open, declarative control plane designed specifically for agent execution. Born from research at Google and Google DeepMind, AX is built to handle agentic workloads that the company describes as a new computing paradigm: stateful, bursty, long-running actors that fluctuate between intense computation and periods of waiting for human, tool, or model responses.

According to agentexecutor.io, traditional orchestrators designed for batch jobs or stateless microservices are often cost-prohibitive for these workloads because they lack native support for sub-second suspend and resume capabilities. AX addresses this by running on Agent Substrate, a compute runtime engineered for high density and fast stateful actor lifecycles. This design enables the platform to support billions of concurrent agent sessions within a single cluster.

AX utilizes four primary primitives to manage these workloads: * **Task**: Provides isolated execution of untrusted code within sandboxes with specific memory and CPU limits. * **Workspace**: Automates the setup of Git repositories, skills, and MCP servers, allowing users to describe environment goals in plain English. * **Gateway**: Manages network policies via explicit allowlists for hosts and ports. * **Model**: Centralizes the configuration of model parameters and secrets.

By utilizing dense multiplexing, AX allows dozens of tasks to share worker resources, ensuring compute costs are only incurred when agents are actively running code or processing. The platform is positioned as a tool for researchers and developers to run reinforcement learning loops, interactive coding agents, and headless browser testing at scale.