We build AI agents and sign our name to them: every one ships with a certificate — evidence attached, expiry date printed.

The story

A century ago, a loom learned to stop.

Not to run faster. Not to weave more. To stop — the instant a single thread broke, before one meter of flawed cloth left the frame. The insight looks small. It is the entire inheritance: a machine that knows when something is wrong is worth more than a machine that merely works.

The factory floor that grew from that loom discovered something stranger. Its highest output was not the thing on the line. It was the person standing beside it, becoming more capable than they were yesterday. Develop the people before you develop the products. The production that produces the producer.

We take this one step further. Our line does not produce cloth or cars. It produces minds. And a mind, like a blade, takes its nature from the profile chosen first. So before the tools, before the skills, before the model is even selected, we write the soul document: this is who you are, this is what you value, this is what you would refuse. Two agents can run on the same model and be fundamentally different beings. The shape is chosen before the steel is folded.

Then the line runs, and the line is honest. Our first assembly took 49 seconds and cost $0.37. Testing broke it immediately: three real bugs, found, fixed, documented in a retro. We published the partial first-run success instead of hiding it, because a factory that hides its defects is not a factory. It is a storefront.

And because a mind's quality is never fixed at the gate — models get swapped, environments drift, context goes stale — nothing that leaves this floor is finished goods. Every certificate of conformance carries an expiry date. Deployment is not exit; it is transfer from a high-frequency inspection zone to a low-frequency one. When the certificate lapses, the unit comes back to the gauges.

The industry tapes a prompt to an API, ships it, forgets it, and starts over. One artisan, one product, every time. That is not innovation. That is the absence of process, a hundred years late. We stand inside an older discipline: standardize the work, build quality in, stop the line, improve every day, respect the people. Then we hand it the newest material there is.

The agents are proof the factory works. The certificates are proof you can sleep.

What we make

Agents, tools, and skills — each assembled from tested, reusable parts, each carrying its own quality checks, each beginning with a soul document.

But that is the inventory, not the product. What this factory actually manufactures is verified claims about agents: trust, in certifiable units. An agent without its certificate is inventory. The certificate — the claim, the evidence bundled behind it, the expiry date printed on it — is what leaves the floor.

Details will live in the knowledge base — coming soon

The promise

This is what a stranger can hold us to. Not what we aspire to — what we can be caught failing at.

Every agent ships with a certificate of conformance. The evidence is attached: what was tested, what was found, what was fixed. If we cannot certify it, we do not ship it. And when the agent you trusted breaks at three in the morning, the certificate says who takes the call.

Every certificate expires. An agent's quality is a claim about the future, and claims about the future decay. The expiry date is our admission of that, and our commitment to re-earn the claim on schedule. We do not sell permanent trust. Nobody honest can.

The soul comes first, in writing. Before the tools, before the model is selected, the soul document says who the agent is, what it values, what it refuses. You can read it. When the agent acts, you can check the act against the document.

Defects go public. Our first production run partially succeeded: three real bugs, caught in testing, fixed, and documented in a retro anyone can read. The retros stay open. Judge us by what we admit, because what a factory admits is the ceiling of what it will ever fix.

Silence counts as failure. An agent that stops producing and says nothing has failed, even if no error fired. In a plant of workers who cannot complain, silence is not health. Our alarms ring for absence, not just for defects.

When something is wrong, the line stops. Not after the sprint. Now. Silent failure is the only unforgivable sin.

Details will live in the knowledge base — coming soon

Principles

Stop the line — jidoka. The loom's lesson, kept whole: quality is built in, not inspected in later. Every agent carries its own checks and surfaces its own failures. And because our workers cannot complain, the alarm rings for silence too — a line that produces nothing is a stopped line, not a calm one.

Pull, not push. We build agents when there is demand for them, not when there is budget for them. No speculative inventory. When the work calls for a specialist, the specialist is shaped — not before. The alternative is creating souls with nothing to do.

Improve daily — kaizen. Every agent better than the one before. Every failure feeds back into the system. The retro is not paperwork; it is how 49 seconds of output made the whole factory smarter. The factory is never finished, and that is its deepest feature; a finished factory is a dead one.

Souls before tools. The soul document is written before the tools, before the skills, before the model is selected. Identity is a stamped structural part — the first line on the bill of materials — not a coat of paint applied at the end. We do not configure agents. We shape them.

Open source. The docs are open. The decision records explain why. The retros show what broke. Everything that defines a unit lives in plain files and git, owned outright, portable by construction. You can burn a library. You cannot burn the act of remembering.

Decentralization. Protocols over bureaucracies. Rules arising from voluntary transactions among free agents, not from a headquarters. What we make belongs to no one and serves everyone; the making happened, and who held the tools is secondary. Trust you can verify does not require trusting us.

Details will live in the knowledge base — coming soon