In plain terms: hrma is a real HR system — payroll, leave, attendance — used by a real company, and roughly seven out of every ten commits that built it were written by an AI agent, with a human reviewing and gating every one before it merged. This note is the starting point for everything else in this garden about how that actually works day to day.
What hrma is
A PDPA-aware (Malaysia’s data protection law) HR platform for Malaysian SMEs, packaged as single-tenant SaaS — one deployment per customer company. The stack: Laravel 12 + Filament for the HR admin panel, a React employee web app, a Capacitor-wrapped mobile app for iOS and Android, and Postgres underneath. It is not a demo — a real company pilots it for real payroll, leave, and attendance, hosted the way described here.
The build, in numbers
145 commits from first scaffold to today, spanning roughly nine weeks. 104 of them — about seven in ten — carry an AI agent’s co-author trailer in the commit message: implementation, fixes, deployment config, and docs, not just boilerplate. Every one of those still went through the same gate a human-written commit would: review before merge.
The repo also carries design documents for individual features — role permissions, attendance geofencing, multi-company support — written before the corresponding code, the same discipline described in write the plan before the code, just applied project-wide rather than to one module map.
What “AI agent built” actually looks like day to day
The number alone (seven in ten commits) undersells the interesting part, which is where humans stayed load-bearing:
- Review: a gap-finder audit compared the API against the admin panel and found a real payroll parity bug — and a human reverted the one agent-proposed fix that was technically correct and organizationally wrong.
- Hardening: an authorization sweep caught managers who could approve claims outside their own team, before any real money moved.
- Maintenance: when a seeder bug broke CI silently, the fix became a standing entry in the repo’s own agent context so the next session can’t reintroduce it.
None of these were the agent working alone, and none were a human reviewing line-by-line either. It was closer to a hand-off: the agent does the exhaustive grinding work — every route, every code path, every CI log — and the human makes the one call at the end that the exhaustive pass can’t: does this decision match how the business actually works.
What I learned
“Built end-to-end with AI agents” is true and also the less interesting half of the story. The more useful fact is the shape of what stayed human: not typing, but judgment — which fix to keep, which access to grant, which failure is allowed to happen twice. Seven in ten commits came from an agent; zero in ten decisions about what “correct” means for this business did.
Related: Letting an AI hunt for the bugs humans skim past, The sweep that found managers could pay other teams’ claims, Teaching AI assistants to remember past mistakes, Hosting a company’s HR system on a PC that never turns off