HOME / 02

THE MACHINE.

Not chatbots. Not demos. Multi-agent systems that run real operations daily — with quality gates, human approval, and the ability to fix their own rulebooks. Here's the shape of it. The details are the moat.

THE ARCHITECTURE// one orchestrator. many specialists. zero trust.

anuj@fleet:~$ cat architecture.txt

             ┌──────────────┐
   inbox ───►│ ORCHESTRATOR │◄─── chat / schedule
             │ (delegates,  │
             │  never works)│
             └──────┬───────┘
        ┌───────────┼───────────────┐
        ▼           ▼               ▼
   [research]   [writers]      [designers]
        │           │               │
        ▼           ▼               ▼
   validators ► independent reviewers (fresh session,
        │       zero producer context — SHIP or REVISE)
        ▼
   HUMAN APPROVAL ── nothing publishes without it
        ▼
   [scheduler] ─► platforms ─► [analytics]
                                   │
        ┌──────────────────────────┘
        ▼
   lessons written INTO the agents' instructions
   (a separate meta-agent patches the rulebook itself)

THE OPERATING RULES// high-level by design

THE ORCHESTRATOR NEVER DOES THE WORKIt reads the task, picks the playbook, and delegates. Heavy lifting happens in specialist agents with their own tools and fresh context.
PRODUCERS NEVER GRADE THEMSELVESEvery deliverable passes an automated validator, then an independent reviewer spawned with zero knowledge of how the work was made. SHIP or REVISE — bounded rounds, then escalate to the human.
HUMANS HOLD THE LAUNCH KEYSNothing reaches a live platform without explicit approval. Automation earns trust by being stoppable.
THE LOOP CLOSESPerformance data flows back and gets distilled into rules the writing and design agents read on their next run. The system learns from what actually worked.
THE RULEBOOK PATCHES ITSELFWhen a mistake happens, a separate maintenance agent locates the responsible instruction file and proposes a surgical fix. Corrections become permanent, not repeated.

IN PRODUCTION

10BRAND ACCOUNTS RUN BY ONE SYSTEM (CLIENT, ANONYMIZED)

13SPECIALIZED SUB-AGENTS, EACH WITH ONE JOB

160+POSTS SHIPPED END-TO-END IN ITS FIRST FIVE WEEKS

And at home: a personal fleet — a design director, a video editor, a web engineer (it built the site you're reading), and a maintainer that keeps the others honest. Same philosophy, smaller blast radius.

The early instincts are public on GitHub: an unofficial Claude API client, an NLP→video Shorts pipeline, TTS experiments in 3D audio — receipts →

Next: 03 — films → · 04 — graphics → · ← 01 apps · ← home