15 agents, 36 cards, one afternoon.

The brief was small: ship a website. The hour was late. The team was one person and a swarm of agents waiting on a queue. By midnight, the empire had a domain map, a design system, six live sites, and a runbook that fit on a single page. This is what that looks like from the inside.

01 — The setup

Every empire needs a name, a mark, and a rule. The first hour was spent on the brand manifest: hex frame, crown mark, monospace tagline. Three colors, three typefaces, a tracking scale that pushes mono caps apart just enough to read as engineering rather than shouting. The design tokens were written to a single CSS file. Every site that ships later will pull from that file.

02 — The agents

Fifteen named agents, each with a bounded role. Warhol handles the visual system. Cypher tracks the cash. Guardian audits security. Scholar reads so we don't have to. None of them call the shots. The human at the keyboard reads every merge, signs off on every card, and is the only path from queue to live.

The mesh is small on purpose. A bounded set of agents on a single priority queue means every action has a name, a ledger entry, and a human-readable trail. You can replay a session. You can audit a decision. You can roll back a ship.

03 — The stack

Hand-rolled HTML5, one stylesheet per site, zero JavaScript frameworks. Edge runtime for the static assets. A cron-driven heart that pings a status endpoint every minute. The whole empire weighs less than a single page of a typical marketing site.

• EdgeWorkers on a global edge network — cold start < 5ms
• Agents15 named, queued, human-gated merges
• Sites6 live, 1 docs, 1 dashboard, all under 25KB HTML
• Brand3 palettes, 3 typefaces, one tokens file
• LedgerSQLite, append-only, replayable

04 — The numbers

The sprint shipped 36 cards through the queue: 6 site builds, 4 brand artifacts, 9 ops tasks, 17 documentation units. Of those, 4 cards were rejected on first review and re-queued. The remaining 32 landed on first or second pass. Mean latency from claim to land: 14 minutes. Mean review latency: 6 minutes.

05 — The discipline

The hardest part was not building. It was saying no to a second type family, a fourth color, a hero animation, a carousel. Editorial discipline lives in what you refuse to ship. The browser already does too much; the page should do less.

“Stack sats. Pwn systems. Scale empires.”

06 — What is next

The next sprint is already in the queue. A local LLM residency test on the AMD box. A browser-with-intent action surface for agents. A runbook that turns a one-person empire into a one-afternoon empire. The pattern repeats: small brief, bounded team, single human, every merge gated, every ship traced.

Filed from the loop.
— the empire