AX – Google’s Open Agentic Orchestrator

summary

AX is a declarative control plane designed specifically for agentic workloads, which differ from microservices and batch jobs due to their stateful, bursty nature. By providing primitives for isolated execution, workspace setup, network security, and model configuration, AX allows developers to run billions of concurrent agent sessions. It leverages Agent Substrate to enable sub-second resumption and dense multiplexing, ensuring cost-efficiency and scalability. This approach removes the infrastructure burden from researchers and developers, letting them focus on agent logic rather than environment management.

key points

1. Agentic workloads require a unique architecture because they alternate between intense computation and idle periods while waiting for model or human responses. AX solves this by checkpointing and suspending idle agents, allowing them to resume in under a second without cold-start delays.
2. The platform uses a declarative system consisting of four primitives—Task, Workspace, Gateway, and Model—to manage untrusted code in sandboxes with strict resource and network limits. This ensures that agents can be deployed safely and consistently across massive clusters.
3. AX integrates generative capabilities directly into its runtime, allowing users to define workspaces using plain English descriptions. The system then automatically prepares the environment and verifies dependencies before the task begins.

who should read it

AI researchers and software engineers building large-scale agentic systems. It is particularly useful for those needing reproducible sandboxes for reinforcement learning or high-density agent deployments.

discussion positive