Zarq
AI orchestration engine that turns rough ideas into merged code — shape, plan, execute, verify, with a human gate at every phase.
Zarq started as a task board with AI agents attached. It’s become the centerpiece of my personal AI infrastructure: an idea-to-merged-code pipeline where I approve shaped ideas, agents on the homelab run them, and I get pinged on Discord when work lands.
The name is a play on “Ark” — as in Noah’s. The doctrine behind it: bring people who don’t write code into working with AI agents, by making every artifact readable at two altitudes (technical detail and plain English).
How the Pipeline Works
Every piece of work moves through phases, each with a specialist and a human gate:
- Shape — A rough idea becomes an aligned brief, task breakdown, and plain-English TLDR
- Plan — The brief becomes an executable plan with dependencies mapped
- Execute — Agents run the plan; work happens in isolated worktrees
- Verify — Independent review before anything counts as done
I stay the director. The agents do the work, but nothing advances a phase without my sign-off.
The Evolution
The original Zarq Board ran task CRUD, agent coordination (Ping and other agents picked up board tasks), automated review, and real-time updates via SSE — all as systemd services on the homelab. It worked, but it was welded to one machine and tried to do too much.
In June 2026 I locked in a reorientation: archive the board, keep the doctrine, and rebuild as a narrow Zarq Engine — a shaping skill plus a minimal Bun CLI that writes Markdown artifacts to a dedicated folder any agent can read. Idea in, aligned brief out, no lock-in to any single runtime.
Why It Matters
Most of my other projects are things Zarq can now build. It’s the difference between me writing code at night and me reviewing shaped, verified work that landed while I was at my day job.