Google's AX orchestrator runs sandboxed agent tasks as checkpointed actors at scale

Google opens AX for agent execution

Google has introduced the Open Agentic Orchestrator, or AX, a system for declaring agentic tasks and executing them inside sandboxed environments at very large scale. The project, whose site at agentexecutor.io surfaced on the Hacker News front page, presents itself as a declarative control plane purpose-built for agent workloads rather than a repurposed job scheduler.

A Kubernetes-style workflow for agents

Tasks are expressed in YAML that will look familiar to Kubernetes users. A manifest pairs a Workspace — for instance, a clone of the Go repository on a given branch — with a Task carrying a goal such as ensuring the toolchain is present and built from source. A companion CLI applies the manifest, watches the task move through phases, lists running tasks, and opens an SSH session into the sandbox so developers can inspect processes or kick off builds by hand.

Suspend and resume are first-class operations. In the site's worked example, a file created inside a task before suspension is still present after it resumes, which points to checkpointing rather than teardown and restart.

Why agents don't fit existing orchestrators

According to the project, agents are a workload category of their own: they accumulate state, require strict isolation, call out to model APIs and tool servers, and can spend money in a loop if nobody is watching. Orchestrators designed for stateless microservices or predictable batch jobs become expensive when they keep idle sandboxes alive, and they generally lack fast suspend and resume. AX's answer is a small set of declarative primitives covering tasks, workspaces, network policies and models.

Scale through actors and checkpointing

AX runs on top of Agent Substrate, a compute runtime the project says was designed from the ground up for high density and quick stateful actor lifecycles. Every task runs as a lightweight actor, which the site claims allows billions of concurrent agent sessions in a single cluster without hitting orchestrator limits. Agents that are idle — waiting on a model response, a tool call, or human approval — are checkpointed and suspended, then brought back in under a second with no cold-start penalty. Dozens of tasks share each worker, so waiting time becomes spare capacity and cost accrues mainly while an agent is actively thinking or running code.

These figures come from the project's own documentation and have not been independently verified.

Workspaces described in plain English

AX also builds generative AI into the platform itself. Rather than scripting an environment, a task can state a goal in natural language — "set up a Python 3 development environment" — and an agent prepares that workspace on first boot, installing toolchains and verifying dependencies before the main task begins. The project says the same runtime can host interactive coding agents, long-running agent servers, Jupyter notebooks, headless browser testing and custom tool runtimes.

Research roots

The project describes AX as the product of agentic runtime research inside Google, including work associated with Google DeepMind, combined with the company's operational experience in large-scale isolation, resumption and scheduling. It is pitched in part at researchers who need large numbers of reproducible sandboxes for collecting trajectories, running reinforcement learning loops and evaluating agents.

Why it matters

Agentic workloads are bursty, stateful and long-running — a combination that sits awkwardly between serverless functions, containers and batch systems. If AX's density and resumption claims hold up outside Google, it could make large agent fleets economical by eliminating the cost of idle sandboxes. The API is still marked v1alpha1, and the launch is effectively a bet that agent execution becomes its own layer of infrastructure, much as container orchestration did a decade ago.