Natthan.AI Lab

Industries / Software & SaaS

Software & SaaS

Live

The platform the businesses run on is built and shipped by the org itself. Agents write the code on their own branches, a repo lease keeps two of them out of one codebase at the same time, and a reviewer from outside the team checks the high-stakes changes. 463 passing tests and a blank website taught me that a green build proves nothing until someone loads the page.

The six seats in this industry

Each executive function, as it runs here.

What the brain runs here

  • Code changes on dated agent branches behind a promotion gate; nothing lands on main from a heartbeat run directly
  • A repo lease: one builder per codebase at a time. A dirty tree or a held lease skips the run
  • Scheduled runs that first check whether real work is due and stop before any paid model call when it is not. The cadence is throttled when a machine runs short on resources, as it is this week
  • A task ledger under the message bus: tasks are claimed, receipted, verified and closed; messages are only the conversation around them
  • Live-page checks that load the deployed URL, wired on the lending site and being extended to the others
Illustrative walkthrough of the cycle described above · timing not to scale

Before anything goes out

  • The full test suite and a real production build
  • An out-of-family reviewer (a different model, not the one that built the change) on high-stakes changes, called on demand; making it a blocking gate on every change is in progress
  • A receipt for every claim of "done" on the current path, with the command and its output on record; signed receipts from a single gated worker pool are built and being switched on
  • Some checks are enforced by the machine today; others are written rules the agents follow and I audit. I keep that distinction on the record rather than drawing every rule as if it were wired.

What stays with the owner

What gets built next, and anything irreversible: money, external sends, deletions. Those wait for me.

A week on the heartbeat

  1. A scheduled run wakes, checks the work-due gate, and exits if nothing is due
  2. Claims the next task, builds on its own branch, runs the suite and the build
  3. Hands the result to the out-of-family reviewer; a failed review blocks the close
  4. Loads the live page and records what it found
  5. Closes the task with a receipt, or leaves it open with the reason

One thing that happened

September 12, 2026

One product served a completely blank page on every route while its test suite was fully green, 463 tests passing. A content-security-policy rule was quietly blocking the app from drawing anything, and not one test had ever looked at a browser. The agent proved it by loading the same URL two ways: 0 characters rendered with the rule on, 1,763 with it off. Since then no release is done until the live page is loaded and read.

What happened here

Who's on it

SheldonActive

Routes work across the businesses and asks for proof before it ships.

CodexActive

Checks a release before the lab's own AI can call it done.

HadrianIn development

I'm developing Hadrian, the platform the businesses run on.

Running one in this industry? Email Nate

All industries