Google has open-sourced AX, a declarative orchestrator specifically engineered to handle the unique demands of agentic workloads. Unlike traditional microservices or batch jobs, AI agents are stateful and bursty—computing intensely for short periods before idling while waiting for model or human responses. AX addresses this by treating agents as lightweight actors, enabling the execution of billions of concurrent sessions per cluster.

Psst!Add AlexTech.ai asPreferred Source on Google

A New Paradigm for Agentic Compute

The system is built on Agent Substrate, a compute runtime designed for massive density and fast stateful actor lifecycles. This architecture allows AX to implement sub-second resumption, where idle agents are checkpointed and suspended, then brought back online almost instantly with zero cold-start delay. To optimize costs, AX uses dense multiplexing, allowing dozens of tasks to share worker resources so compute is only paid for when agents are actively processing.

Declarative Primitives and Generative Setup

AX simplifies agent infrastructure through four core primitives: Tasks for isolated execution in sandboxes with CPU/memory limits; Workspaces for defining Git repos and MCP servers; Gateways for managing network allowlists and credentials; and Models for centralized configuration of parameters and secrets.

A standout feature is the integration of generative AI directly into the platform. Users can describe a required environment in plain English, and AX will automatically prepare the workspace by handing the goal to an agent during the first boot to install toolchains and verify dependencies.

Evolution Toward Production Scale

While previous industry trends showed a convergence toward standardized agent architectures, as analyzed in the shift toward corporate agent operating systems, AX represents a move toward high-throughput production. According to Ai Tldr, the project has recently evolved in version v0.3.0, splitting into three services and migrating task state management from Kubernetes to Redis for improved performance.

This infrastructure is particularly suited for AI research, allowing the creation of massive numbers of reproducible sandboxes to collect trajectories and run reinforcement learning loops at scale. The project is now available on GitHub.