AX: Google's open orchestrator for agentic tasks

AX is an open-source control plane for running AI agents: a task is declared in YAML, and the system sandboxes it, wires up its workspace, fences its network and helps run large numbers of them per cluster. The project is published on GitHub as google/ax under the Apache 2.0 licence and presented at agentexecutor.io.

Four primitives

AX reduces agent infrastructure to four declarative objects. A Task runs untrusted agent code in a sandbox with CPU and memory limits and is cheap to create, suspend or throw away. A Workspace lists the Git repositories, MCP servers and skills an agent needs — or simply describes the goal — and is prepared in every sandbox before the task starts. A Gateway defines network policies, locking traffic to an explicit allowlist of hosts and ports and injecting credentials into outgoing requests. A Model block keeps models, parameters and secrets in one place, so rotating a key or pinning a model version is one apply away.

How it runs

AX runs on a runtime the project calls Agent Substrate, built for density and fast stateful actor lifecycles. Each task runs as a lightweight actor, which the project says allows scaling to billions of concurrent agent sessions per cluster without orchestrator limits. Idle agents waiting on model responses, tool calls or human approval are checkpointed and suspended, then resumed in under a second with no cold start, and dozens of tasks share worker resources so idle waiting time becomes spare compute. The command line follows the same declarative model: ax apply -f task.yaml creates the workspace and the task, ax watch and ax get tasks report the phase and worker IP, ax ssh opens a shell inside the sandbox, and ax suspend, ax resume and ax delete manage its lifecycle.

Where it came from

The project describes itself as born at Google, where research on agentic runtimes met frontier compute, with work drawing on Google DeepMind. Its argument is that agents are neither microservices nor batch jobs: they accumulate state, need strict isolation, call out to model APIs and tool servers, and waste resources if nobody is watching them. Traditional orchestrators built for stateless services become cost-prohibitive when they keep idle sandboxes running, and lack native sub-second suspend and resume. AX is also aimed at researchers, who can spin up reproducible sandboxes to collect trajectories, run reinforcement learning loops and evaluate agents at scale.

Generative features are part of the platform as well: a workspace can be described in plain English, and AX hands that goal to an agent on first boot to install toolchains and verify dependencies. According to the project, it runs interactive coding agents, long-running agent servers, Jupyter notebooks, headless browser testing and custom tool runtimes.