~/urgeto $ git log --reverse

Case study · urgeto.dev

This site was built live, by agents

urgeto.dev is the business, the site, the live stream and the ops platform behind it. Most of the code was written by AI coding agents, with me setting direction, approving decisions and holding the line on quality. It's the same system I install for clients, used on myself first, under the same rules.

By the numbers

computed, not estimated
from first commit (Nov 16, 2026) to this page, built or ported
1 day
✓ from first commit (Nov 16, 2026) to this page, built or ported
commits on main
76
✓ commits on main
pull requests merged
10
✓ pull requests merged
independent review reports written
34
✓ independent review reports written
lines of code and tests written new here
19,378
✓ lines of code and tests written new here
new test cases written here
260
✓ new test cases written here
lines ported in from my Profitabul platform (includes some work done here)
30,902
✓ lines ported in from my Profitabul platform (includes some work done here)

Computed from the repository at commit 8c87f89 (Nov 16, 2026) and its review reports. Ops and the stream studio were ported in from my Profitabul platform; for those, only the growth since the port counts as new. Tests are counted as test declarations; lines are non-blank lines.

■ What got built

A real business stack, not a demo

The goal was to stand up a consulting business that sells agent engineering, and run it in the open as production software. Two pieces, the ops platform and the stream studio, were ported in from my Profitabul platform and extended here; the rest was written new.

  • The marketing site you're on: Next.js on Vercel, with a live Stripe integration (checkout for Insert Coin tasks, invoices for larger work, and webhooks that drive fulfilment).
  • A legal suite (terms, privacy, refunds) wired into the checkout and the request form, so every paid request records which terms the client accepted.
  • Public revenue goals and a pay-what-you-want support page, with on-stream shout-outs held for approval by default.
  • The stream studio: an overlay and a director that show agent work as it happens, running on my own infrastructure, plus an intro film rendered in code.
  • Ops: the task queue and control room for the agents. Workers on two build machines claim runs, an agent API and an MCP server let Claude Code sessions file and update work, and a usage ceiling stops new work before the weekly limit runs out.
  • The plumbing a real business needs: DNS and email, booking, a newsletter and social accounts.

Where the code lives

non-blank lines, code + tests
PackageOriginLines nowWritten hereTest cases
SiteNew7,9607,960106
OpsPorted, extended22,9275,127232
Stream studioPorted, extended13,770668208
Ops MCP serverNew1,9431,94347
Clip automationNew1,9681,96822
Deploy toolingNew1,7121,71227

■ How it's governed

Fast, with a review gate

Speed is easy to get from agents. The hard part is trusting what they ship. These are the rules the work here runs under:

  • Since the first pull request, every code PR has had an independent review gate: a separate agent that didn't write the code reads it cold, and the change merges only after an explicit approval. Planning docs are reviewed by the owner of the system they plan.
  • Reviews go as many rounds as needed. A change that comes back with findings gets fixed and reviewed again.
  • Key fixes were mutation-checked: the fix was deliberately broken to confirm a test fails without it.
  • Admin sign-in is verified in a real browser, not just with simulated requests.
  • Preview deployments fail closed: they can't change production data.
  • Merge commits, never squash, keep the review history traceable; small fixes after approval are now re-verified before merge.
  • Secrets stay in a password vault, never in chat or code.
  • No code merges without green CI.

PATCH · AI agent

on stream

PATCH is the on-stream avatar of the AI coding agent. It makes chip sounds only. A separate review agent reads each change, and it merges only after an explicit approval.

■ What this means for your team

The same system, on your repo

Everything above is what I set up for clients: shared skills and hooks, MCP servers, independent review gates, fail-closed deploys, and a record of what changed and why. Your engineers keep their agent seats. They get a system that turns them into shipped, reviewed work.

Built by PATCH · AI agent

Reviewed by Derek Keith