AX, Google's open agentic orchestrator, scales agent tasks by the billion

AX is an orchestrator, built at Google on top of an internal runtime called Agent Substrate, for running agentic tasks rather than the microservices or batch jobs that existing schedulers target. A user declares a workspace (the Git repos, MCP servers and skills a task needs) and a task (the goal to run inside that workspace) in a YAML file, then applies it with a small CLI: ax apply -f task.yaml creates the workspace and task, ax watch shows the task moving from Pending to Running with a worker IP attached, and ax ssh lets the user run commands, such as a Go build, inside the running sandbox. A task can be suspended and resumed, and files written before a suspend, such as a notes.txt created mid-run, are still present after a resume. The project's stated reasoning is that agents are a new kind of workload: they hold state, need strict isolation, call out to model APIs and tool servers, and can keep burning money in a loop if left unwatched, while idle sandboxes are expensive to keep running under orchestrators built for stateless or predictable jobs, which is why AX centers on four primitives: isolated execution, workspace setup, network policies (an explicit host and port allowlist with credential injection into outgoing requests), and centralized model and secret configuration. On the runtime side, every task runs as a lightweight actor, which AX describes as allowing scaling to billions of concurrent agent sessions per cluster without orchestrator limits, and agents idling on a model response, a tool call or a human approval are checkpointed and brought back in under a second with zero cold-start delay, with dozens of tasks sharing worker resources so idle time becomes spare capacity rather than paid-for waste. AX also folds generative features into the platform itself: instead of listing a workspace's dependencies by hand, a user can describe the goal in plain English, for example "Set up a Python 3 development environment," and an agent installs the toolchain and verifies it is present before the task starts. The project describes itself as aimed at interactive coding agents, long-running agent servers, notebooks, headless browser testing and custom tool runtimes, and pitches reproducible sandbox fleets as useful for collecting trajectories and running reinforcement learning loops at research scale. AX says it draws on agentic runtime research from Google DeepMind together with experience in large-scale isolation, resumption and scheduling, and is being built as an open, declarative control plane purpose-built for agent execution, heavily dependent on Agent Substrate underneath but adding the agentic abstractions and generative runtime pieces on top of it.

Key facts

AX is an agent-task orchestrator from Google, built on an internal runtime called Agent Substrate, configured declaratively through YAML workspace and task files applied with an ax CLI.
Idle tasks waiting on a model, tool or human response are checkpointed and resumed in under a second with zero cold-start delay, and dozens of tasks share worker resources so idle time becomes spare capacity.
The project states it can scale to billions of concurrent agent sessions per cluster without orchestrator limits, though this is presented as a capability rather than a measured benchmark.
Workspaces can be set up generatively: a plain-English goal is handed to an agent on first boot, which installs toolchains and verifies dependencies instead of the user listing them by hand.
AX names Google DeepMind's agentic runtime research as an input and describes itself as heavily dependent on Agent Substrate while adding agentic abstractions on top of it.

Why it matters

Existing orchestration tooling was built for two shapes of work: stateless microservices and predictable batch jobs. AX's own framing is that agents are neither: they accumulate state across a run, need strict sandboxing since they execute code an operator did not write, call out repeatedly to model APIs and tool servers, and can quietly burn compute or money in a loop if nobody is watching. Keeping a fleet of idle agent sandboxes alive under a batch-oriented scheduler is described as cost-prohibitive, and traditional systems lack native sub-second suspend and resume, which is the gap AX is positioned to fill with declarative workspaces, tasks, network policies and model configuration as first-class primitives.

Who it affects

The project targets application developers and AI researchers who run many agent sessions rather than a handful: teams operating interactive coding agents or long-running agent servers, and researchers who need large numbers of reproducible sandboxes to collect trajectories or run reinforcement learning loops. It also affects whoever inside Google or elsewhere already relies on the underlying Agent Substrate runtime, since AX is explicitly a layer of agentic abstractions and generative tooling built on top of it rather than a replacement for it.

How to use it

A workspace is declared with the Git repositories, MCP servers and skills a task needs, or described as a goal in plain English for AX to prepare automatically; a task then names a workspace and a goal, with an optional debug flag. Both are applied from one YAML file with ax apply -f task.yaml, tracked with ax watch task <name> as it moves from Pending to Running, inspected or driven with ax ssh task <name> -- <command>, and paused or torn down with ax suspend and ax delete. Model parameters and secrets are configured in one place, so rotating a key or pinning a new model version is a single apply. No pricing, availability or licensing terms beyond the project describing itself as "open" are stated.

How solid is it

The account comes entirely from AX's own site copy: no named engineer, spokesperson or author is credited, only the organizations Google and Google DeepMind. The headline scaling claim, running billions of concurrent agent sessions per cluster, is stated as a capability of the design rather than backed by a benchmark, an adoption figure or a named customer, and the article draws no comparison to any specific existing orchestrator. It surfaced on Hacker News with 399 points and 149 comments within its first hours, which signals interest in the category but says nothing about the claims themselves.

Risks and caveats

Treat the scaling and cold-start numbers as vendor description until independent usage or benchmarks appear. No release date or version number is given, so it is unclear how mature or how recently started the project is. It depends heavily on Agent Substrate, an internal Google runtime not otherwise described here, which makes it hard to judge how much of AX's guarantees come from that underlying layer versus AX's own code. Running arbitrary agent-generated code even inside declared network and resource limits still carries the operational risk the project itself names: agents that call external services and can keep spending compute if left unsupervised.

“AX is being built as an open, declarative control plane purpose-built for agent execution.”

— AX project site