Google open-sources AX, a Kubernetes-style orchestrator for agents (InfoQ)

InfoQ news item by Olimpiu Pop, 22 September 2026, on AX: an Apache-2.0 orchestrator from Google that runs autonomous agent workloads on a Kubernetes cluster, on top of Agent Substrate. Arrived by Telegram on 2026-09-23.

T3, following this spoke's standing InfoQ posture: the article reports and adds community reaction, but measures nothing. This page checked its technical claims against the AX repository (README, DESIGN.md, docs) and the Agent Substrate README the same day. The facts below come from those primaries unless marked as InfoQ's.

The argument

Agents are a third kind of workload. They are not stateless services and not run-to-completion batch jobs. They hold state, run for a long time, and sit idle for most of it while they wait on a model API, a tool, or a human. On an ordinary cluster an idle agent keeps its sandbox, so compute sits unused; a cold start on resume slows an interactive loop. AX's answer is to treat each agent session as an actor that can be checkpointed and suspended when idle, then resumed on any free worker.

What AX is (from the repository)

Four declarative kinds under ax.io/v1alpha1: Task (the sandboxed run, with CPU and memory limits), Workspace (Git repos, MCP servers and skill packages set up before the task starts, or a natural-language goal an init agent carries out), Gateway (outbound traffic limited to a host allowlist, with credentials injected), and Model (which LLM the platform itself uses, with keys from a Kubernetes secret).
A kubectl-shaped CLI written in Go: apply, get, watch, ssh, suspend, resume.
State lives in Redis, not etcd. DESIGN.md gives the reason: storing "millions of short-lived tasks as Kubernetes CRDs pushes etcd past its comfort zone (single-digit GB storage limits, write-rate bottlenecks, control plane degradation)". A Redis Stream is the work queue between a stateless API server and a pool of reconciliation workers that scale by adding replicas.
Pre-stable. The README warns of "major breaking changes prior to a stable release".

InfoQ's summary of the Hacker News thread lists "custom CRDs" among the operating costs. DESIGN.md says the opposite about task state: it is kept out of Kubernetes CRDs on purpose. This page could not check whether the install creates any CRDs for other purposes, so the complaint is recorded but not confirmed.

The numbers are Agent Substrate's, and they are first-party

InfoQ repeats "sub-second" resume and "zero cold-start delay". The source for those figures is the Agent Substrate README, not AX: sub-500 ms resume, over 500 suspend/resume activations per second, 10x higher density than standard container runtimes, and a demo of about 250 actors on 8 pods. Substrate runs sandboxes on gVisor or microVMs. Its README also says it "is not an officially supported Google product". None of these numbers has been measured by anyone else.

Community reaction (InfoQ's summary; threads not read here)

Infrastructure engineers welcomed a fix for paying for idle agents.
Other developers said the "ergonomic" claim ignores the cost of running a cluster, a registry and ko builds.
Reddit posters placed AX below agent frameworks such as LangGraph or CrewAI: it runs agents, it does not build them.
Early reports mention dropped connections in the egress proxy and basic secrets handling.

Weaknesses recorded

InfoQ's line that AX comes from "systems research from teams across Google and Google DeepMind" does not appear in the repository pages read for this summary.
The article gives no adoption figures, and nobody quoted names a production deployment.
The two diagrams are marked as "Generated with Gemini", so they are illustrations, not architecture.

google-ax · agent-substrate · compute-sandbox · modal-1m-sandboxes · kubernetes · infoq · cloud-run-sandboxes · agent-guardrails