Google's AX Promises Billions of Agents, Demos 250

Google released AX, an open-source runtime for running AI agent workloads at scale, on Sept. 20, and within a day developers on Hacker News were pointing out that its quickstart bore little resemblance to its pitch.

The project, hosted at github.com/google/ax under an Apache 2.0 license, describes itself in its own repository as "a high-throughput, declarative orchestrator to run billions of autonomous agent workloads in a cluster." Google's design document frames the problem as one of idle time: agents spend most of their life waiting on a model reply, a tool call or a human reviewer, and AX is built to suspend and resume that wait in under a second rather than pay to keep a container running while nothing happens.

What the quickstart actually asks for

AX's own homepage, agentexecutor.io, says the project has "an uncompromising focus on ergonomics, rapid iteration, and joyful workflows." Its GitHub quickstart, as commenter alembic_fumes wrote in the top-voted reply on the Sept. 20 launch thread, asks for something else first: "a Kubernetes cluster, ko (brew install ko), a container registry your cluster can pull from, and a reachable Agent Substrate Control API." The thread drew 661 points and 299 comments by Sept. 23. "Call me old-fashioned but I don't find this 'easier,'" the commenter wrote, adding there is "a vast chasm between what this tool is being sold as and what it actually is."

The complaint echoes one this site has tracked on the coding-agent side of the same problem: Claude Code shipped AGENTS.md support that silently failed on exactly the setups, like disabled telemetry, that cautious infrastructure teams tend to run.

Nine thousand stars, one demo of 250

AX's own launch demo, described in the Hacker News thread and in InfoQ's write-up published Sept. 22, ran roughly 250 stateful agent sessions across eight pods, with sub-second suspend and resume between them. That is the only concurrency figure Google has actually shown working, against a repository description that promises billions per cluster. A Sept. 21 account on Dev.to put the repository at 1,968 stars and 623 commits the day after launch; GitHub's own count showed 9,052 stars and 434 forks when this was checked on Sept. 24, after the trending page recorded 1,543 new stars in the most recent day alone. "Until then, 'billions' is marketing," the Dev.to piece concluded, noting no audited benchmark backs the larger figure.

1. Sept. 21, 20261968 stars
2. Sept. 24, 20269052 stars

Source: Dev.to (Sept. 21, 2026) and GitHub API (accessed Sept. 24, 2026)

Redis instead of Kubernetes' own storage

The design document explains one real engineering choice behind the pitch. AX does not store agent tasks as Kubernetes custom resources, because, in the project's own words, "storing millions of short-lived tasks as Kubernetes CRDs pushes etcd past its comfort zone." Instead, a stateless ax-server writes task state to Redis Streams, and a horizontally scaled pool of ax-controller workers reads from that queue to provision sandboxes on Google's separate Agent Substrate project, which isolates each task inside a gVisor kernel.

What AX promises	What the quickstart requires first
"Uncompromising focus on ergonomics"	A Kubernetes cluster
"Rapid iteration, joyful workflows"	The ko build tool and a container registry
Billions of agent workloads per cluster	A reachable Agent Substrate Control API

What Google has not shipped yet

The repository's own README warns that the team is "still actively refining our core concepts, protocols, and specifications" and expects "major breaking changes prior to a stable release." Agent Substrate, the sandboxing layer AX depends on, is itself still in beta, a combination one commenter on the launch thread called running "beta on another beta." Other commenters raised Google's record of retiring developer-facing projects, naming Google Wave, Google+ and Google Reader as prior cases, a skepticism this site has seen aimed at a smaller framework too, when Superpowers drew scrutiny over token costs rather than staying power.

Google has not published a target date for a stable 1.0 release of either AX or Agent Substrate. Until one ships, the 9,052 stars measure attention, not the billions-per-cluster claim the project was built to prove.