Google open-sources AX, an orchestrator that treats agents as actors
Serguey Asael Shinder2 min readGoogle has released AX, an Apache 2.0-licensed orchestrator and declarative runtime for autonomous AI agent workloads, hosted at agentexecutor.io and on GitHub as google/ax. InfoQ reported it on 22 September 2026.
The design claim is that agents are the wrong shape for existing orchestration. A stateless microservice handles a short request-response lifecycle; a batch job runs deterministically to completion. An autonomous agent is stateful, bursty and long-running — heavy compute during reasoning, tool execution and local code evaluation, interleaved with long idle periods waiting on model responses, external APIs or a human. On conventional container orchestration, keeping sandboxes alive through those idle phases wastes compute, while cold starts hurt interactive loops.
AX runs on Agent Substrate, an execution runtime built for dense actor multiplexing, with systems research from teams across Google and Google DeepMind behind it. Each agent session runs as an isolated actor sandbox with strict CPU and memory boundaries. When an agent goes idle, the platform checkpoints its execution state and suspends it, and is designed to resume suspended actors in sub-second intervals with zero cold-start delay, multiplexing dozens of tasks onto shared host workers.
Four declarative primitives are exposed: Task, Workspace, Gateway and Model.
What it means
The interesting engineering claim is the checkpoint, not the primitives. Suspending a running process mid-execution and resuming it in under a second is the hard part, and it is what decides whether the economics work — an agent that spends most of its wall-clock time waiting is only expensive if you pay for the waiting.
Read "Kubernetes style" as a description of the interface rather than the substrate. Declarative objects reconciled toward a desired state is a familiar shape and a genuinely good fit for long-lived work. But the thing underneath is an actor runtime, not a container scheduler, and the operational knowledge that transfers is the vocabulary rather than the runbook.
Two things this announcement does not settle. Apache 2.0 and a GitHub repository say the code is available; they say nothing about whether the substrate is practical to operate outside Google's environment. And "sub-second resumption" is a design goal as reported, not a benchmark we have measured — the number to look for is resumption latency against agent state size, which is where checkpoint systems usually become interesting.