AX wants to run 'billions' of AI agents per cluster - but is it actually Google's?

A new tool called AX is making the rounds this week, promising to solve a problem that barely existed two years ago: how do you run millions of semi-autonomous AI agents without them going rogue, burning through your cloud budget, or falling over the moment they need to pause and wait for a response. The Hacker News crowd has piled in, with nearly 400 upvotes and 149 comments at time of writing. The catch is that some of the most interesting discussion isn’t about what AX does - it’s about who actually made it.

What AX actually does

Based on its own documentation, AX is a command-line tool and runtime for declaring “agentic tasks” - jobs where an AI model needs to read code, run commands, call external tools, and keep some memory of what it’s doing along the way. You write a YAML file describing a workspace (say, a Git repo) and a goal, run ax apply, and the platform spins up an isolated sandbox with network rules and credentials pre-wired.

The pitch rests on four building blocks: Task (an isolated sandbox with CPU and memory limits), Workspace (automatic environment setup from a repo list or plain-English description), Gateway (network allowlisting so an agent can’t phone home to anywhere it likes) and Model (centralised configuration for API keys and model versions). Under the bonnet, AX says it runs on something called “Agent Substrate”, a compute layer built specifically to let idle agents - the ones sitting around waiting for a model reply - be checkpointed and resumed in under a second rather than eating compute the whole time.

That’s the known part. It’s a genuine, working piece of infrastructure with a real CLI and a real GitHub repo, aimed squarely at teams running AI coding agents or automation pipelines at scale.

The claims worth squinting at

Everything past that is vendor marketing, not independent verification. AX’s site claims it can handle “billions of tasks per cluster” and “sub-second resumption with zero cold-start delay.” Those are impressive numbers, but they’re the company’s own framing, tested nowhere that’s been published for outside scrutiny. Nothing in the source material shows benchmarks, customer names, or third-party numbers.

There’s also the small matter of the headline claim doing the rounds on Hacker News: that this is “Google’s” project. Google’s name doesn’t actually appear in AX’s own site copy as quoted in the source material - the “Google” label comes from the submission title, and commenters spent a chunk of the thread debating exactly that point, half-jokingly wondering whether it’s officially Google-backed or just Google-adjacent, and whether it’ll still exist in five years given the company’s history of retiring tools (Google Reader, Google+, and the recently reshuffled Gemini CLI all got a mention).

So who is actually affected

If you’re not building or operating fleets of AI agents, this changes precisely nothing about your day. It’s infrastructure plumbing aimed at developers and platform teams, not consumers. Even within that audience, Hacker News commenters pointed out the market is already flooded with rival agent-sandboxing frameworks, so AX is one contender among many rather than an obvious default choice.

The takeaway

AX is a real, working tool with a sensible design for a genuine problem - agents that idle, misbehave, or need tight sandboxing. But its headline-grabbing scale claims are unverified marketing, and its “Google” branding is murkier than the viral headline suggests. Worth watching if you work in this space; not worth panicking or celebrating about if you don’t.