Google's Open Agentic Orchestrator for AI Workloads
AX platform enables scalable execution of stateful AI agents with declarative infrastructure
1New computing paradigm
AX addresses the unique requirements of agentic workloads that differ from traditional microservices or batch jobs. These stateful actors maintain context, require strict isolation, interact with model APIs, and can incur significant costs if left unchecked. The platform provides four core primitives to declaratively manage these characteristics, representing a fundamental shift in how we approach AI workload orchestration.
2Sandboxed execution
Each task runs in a lightweight, isolated sandbox with controlled CPU and memory allocation. The system enables rapid creation, suspension, and disposal of these environments. This architecture supports running billions of concurrent agent sessions per cluster without hitting orchestrator limitations, while maintaining security through strict resource boundaries.
3Declarative environment setup
Developers can specify required Git repositories, MCP servers, and necessary skills, or simply describe the end goal in plain English. AX automatically configures each sandbox with these requirements before task execution begins. This includes generative AI capabilities that can interpret natural language descriptions to prepare appropriate development environments.
4Secure networking controls
The platform provides granular network policy management, allowing explicit allowlists of hosts and ports. It handles credential injection for incoming requests and supports rapid policy updates. These features address the security challenges of agentic systems that need to interact with external services while maintaining isolation.
5Efficient resource utilization
AX implements checkpointing and sub-second suspend/resume capabilities for idle agents waiting on model responses or external calls. Dozens of tasks share worker resources, converting idle wait time into available capacity. This architecture ensures users only pay for active computation periods while maintaining instant availability.
6Centralized model management
The platform consolidates configuration of AI models, their parameters, and associated secrets in one location. This enables single-command updates like key rotation or model version pinning across all running agents. Such centralized control is critical when managing large fleets of AI agents with shared dependencies.
7Research origins
AX emerged from Google's agentic runtime systems research, combining DeepMind's work with large-scale isolation and scheduling expertise. It represents years of operational experience with agent execution engines, recognizing these workloads as a distinct computing paradigm requiring specialized infrastructure beyond traditional orchestrators.
8Developer experience focus
The platform prioritizes ergonomic workflows and rapid iteration for both application developers and AI researchers. By abstracting infrastructure concerns into declarative primitives, AX enables teams to focus on agent behavior rather than underlying plumbing, accelerating development cycles for agentic applications.
Original article: Hacker News — Google's Open Agentic Orchestrator for AI Workloads