Google releases AX, an open orchestrator for AI agents

Google has released AX, an open orchestrator built for running AI agents. Its premise is that agents are an entirely new class of workload: neither microservices nor batch jobs. An agent accumulates state, needs strict isolation, calls out to model APIs and external tools — and, if nobody is watching, can burn money in a loop. Traditional orchestrators designed for stateless services or predictable batches turn out to be either expensive or ill-suited here.

Four primitives

AX offers a declarative model of four objects described in YAML, in Kubernetes style. Task is the isolated execution of untrusted agent code in a sandbox with CPU and memory limits; such a task is cheap to create, suspend and throw away. Workspace describes the environment: the Git repositories, MCP servers and skills an agent needs. Gateway handles network policies, locking traffic down to an explicit allowlist of hosts and ports while injecting credentials into outgoing requests. Model gathers model configuration, parameters and secrets in one place, so a key or a pinned model version can be changed with a single apply.

Scale and cost

Under the hood AX runs on top of Agent Substrate, a runtime the developers say was designed for massive density and fast stateful-actor lifecycles. Every task runs as a lightweight actor, which lets the project claim billions of concurrent agent sessions per cluster. Idle agents waiting on a model, a tool or a human are checkpointed and brought back in under a second with no cold start. Dozens of tasks are multiplexed onto shared worker resources — you pay, the pitch goes, only when an agent is actually thinking or running code.

A "generative" trait of the platform is highlighted separately: an environment can be described in plain English, and on first boot an agent installs the needed toolchains and verifies dependencies itself.

Where it comes from

AX, the project says, was born inside Google where research on agentic runtimes met frontier compute; it draws on Google DeepMind work in isolation, resumption and scheduling. It is positioned as a friendly runtime for developers and researchers, with promises of a minimal core and an emphasis on reproducible sandboxes for collecting trajectories, running reinforcement-learning loops and evaluating agents at scale.

Source: Hacker News