Google releases open-source declarative AI agent runtime AX
Google has introduced AX, an open declarative control plane purpose-built for large-scale agentic AI workloads. Emerging from research initiatives at Google DeepMind and Google infrastructure teams, AX addresses a fundamental gap in modern AI deployment: the inability of traditional orchestration systems to efficiently handle the stateful, bursty, and long-running nature of autonomous agents. Conventional schedulers, optimized for stateless microservices or predictable batch processing, struggle with agent workloads that alternate between intense computation and extended waiting periods for model responses, tool executions, or human oversight. Maintaining idle sandboxes to bridge these gaps proves economically inefficient, while existing platforms lack native support for the sub-second suspend and resume capabilities required to maintain agent state across interruptions. AX was engineered to resolve these constraints by abstracting agentic execution into a unified control plane. The system enables developers and researchers to declare tasks declaratively, automatically provisioning isolated workspaces, network policies, and model connections. AX can run billions of tasks per cluster, supporting both single-task agents and complex multi-task compositions. Underlying the platform is Agent Substrate, which provides the low-level runtime capabilities, while AX layers high-level abstractions and generative runtime components on top. This architecture ensures secure isolation without sacrificing operational speed or resource efficiency. By decoupling task orchestration from infrastructure management, AX eliminates the need for teams to rebuild foundational agent execution pipelines. The platform is designed to transition seamlessly from research environments to production deployments, offering predictable scaling and cost-effective resource utilization. Early adopters can integrate AX into existing AI development workflows to deploy massive fleets of autonomous agents while maintaining strict governance over workspace access and network boundaries. The release marks a strategic shift toward infrastructure optimized for the next generation of AI applications, positioning AX as a foundational layer for scalable, production-grade agentic systems.
