It started as agents with real jobs.

Before minarray had a name, its agents had jobs: keeping versioned backups of a store, watching a storefront and staying silent while it was healthy, turning a market brief into researched leads. Operating them for real work taught us the thing this company is built on: the question that decides whether an agent can be handed a real job is not what can it do? It is what can it be trusted not to do?

An agent that writes to the store it backs up is a liability, whatever its demo looks like. An agent that restores a theme on a guess is worse than one that pages a person. The boundaries were always the hard part.

So minarray is built as a ledger before it is a product. Every run is scoped to one registered capability, counted against a daily quota, and written to a durable record, including the runs that failed, and the things a run was not allowed to touch. What we can put on this website is exactly what those records support, which is why the registry on the front page says MODELED, in plain sight, seven times.

What it wants to be is the version of this that stays honest at scale: a team you run (talk to each agent privately, bring them together, hand work between them), whose references are its own run records, and you can read them. A small array whose every claim you can check.

A registry you can read in one screen is a scope you can actually promise.

Ten capabilities, three of them live, is small enough that nobody has to guess what minarray does, including us. The registry grows only when the next capability is actually built. The boundaries that make that promise enforceable are documented, mechanism by mechanism, on Boundaries.