Twenty years in enterprise technology, and I still ship the code.

An agent that administers nine servers. It proposes changes to its own capabilities, and it cannot merge them. I do. Every enterprise is going to ask you how you let an AI near production, and this is the answer I have actually run rather than the one I would put on a slide.

For engineers. Nine declarative NixOS hosts, a Proxmox cluster, an OPNsense edge, a self-hosted mesh. The agent develops its own skills and hooks on a branch; only a human merges to the live one. Thirty six of the sixty six architecture decisions were later superseded, which means I was wrong more than half the time and wrote down why on each occasion. That record is the thing an enterprise security review actually asks for. The honest limit: nine hosts in a house, one user, and nobody has ever been angry with me about it at 3am. The scale is not enterprise. The shape of the gate is.

It finds companies that are about to need what you sell, and shows its working. It never contacts anyone. It suggests, and a human decides.

For engineers. I planted fabricated evidence in fifteen cases and measured whether the judge caught it. With a verification tool it caught all of them. Without one it waved two thirds of them through. That second number is why the tool is not optional, and it is the only reason this system is worth anything. The honest limit: fifteen cases is a smoke test, not a benchmark, and I wrote both the cases and the rubric, so the judge and I share a blind spot by construction. A second labeller would be the real test and I have not run one.

It probes six applicant tracking systems directly, pulls every new role, and judges each one against what I am after. A guard drops any fact the model cannot cite. I built that guard after the model invented investor lists for three companies, and four rows in my database still predate it.

For engineers. The rule is: only a probe or a citation may confirm a fact, never the model. It is not a slogan. It is a function that drops any field whose source URL is not in the retrieved corpus, and it has tests, including one that plants a fabricated source URL. I wrote it after the model copied “Sequoia, Accel, Y Combinator” out of an example into three unrelated companies. The guard shipped after the data, so four rows still carry facts it would now reject. They are still there. The honest limit: 39 percent line coverage. Ninety one percent on the ATS parsers, eighteen percent on the sweep that actually runs them. I know which half is load bearing, and I know which half I tested.