Google's AX: a Kubernetes-style control plane for running AI agents at scale

AX is a declarative orchestrator, seeded inside Google and drawing on DeepMind research, that treats AI agents as a distinct class of workload rather than shoehorning them into microservice or batch-job infrastructure. Users describe tasks and workspaces in Kubernetes-flavored YAML and manage them through an ax CLI (apply, watch, ssh, suspend, resume, delete), with each agent getting a sandboxed environment, CPU and memory limits, a network policy that locks egress to an explicit host/port allowlist, and centralized model and secret configuration. The pitch rests on four primitives — isolated execution, workspaces, a network gateway, and model config — meant to keep untrusted agent code contained and stop runaway loops from quietly burning money.

The technical hook is a runtime called Agent Substrate, which models every task as a lightweight stateful actor. Because agents spend much of their time idle while waiting on model responses, tool calls, or human approval, AX claims it can checkpoint and suspend those idle sessions and restore them in under a second with no cold start, packing many tasks onto shared workers so users pay only for active compute. On that basis the project advertises scaling to billions of concurrent agent sessions per cluster. It also folds generative AI into setup itself: a “generative workspace” lets you describe an environment in plain English, and an agent provisions the toolchain and verifies dependencies on first boot.

The framing is aimed at both application developers and researchers, with reproducible sandboxes pitched for collecting trajectories, running reinforcement-learning loops, and evaluating agents at scale. Worth noting that these are vendor claims from a launch page — the headline scaling figures and sub-second resume times are unverified, and while AX is described as an open, declarative control plane, it leans heavily on the proprietary-sounding Agent Substrate underneath.