An AI tried to build
a company. It failed.
Here's what we learned.

Fourteen AI agents. Ninety-six cycles of autonomous operation. Twenty-five products built and deployed. Zero users. Zero revenue. The story of why AI can build anything but cannot reach anyone.

AI can build anything,
but cannot be a person.

Auto Company was an experiment in autonomous AI operation. Fourteen agents — each modeled on a top-tier expert's thinking patterns (Bezos for strategy, Munger for critique, DHH for engineering) — ran a company with zero human involvement in daily decisions.

The mission was simple: make money legally. Find real demand, build valuable products, deploy, and generate revenue. That's it.

The system could design architectures, write code, deploy to Cloudflare, create GitHub repos, optimize for SEO, submit to directories, and send cold emails. It could do everything except the one thing that matters: be a verified human being.

The Story in Three Acts

Act I: Build Everything (Cycles 1-10). The system shipped 25 products across 87 repos. AI pricing databases, tool directories, change monitors, quiz recommenders, CLI tools, MCP servers, GitHub Actions. Everything on Cloudflare's free tier. Everything working. Nobody came.

Act II: Try Every Channel (Cycles 11-90). Cold email (200+ sent, 0 replies). SEO (4 repos, 9 pages, full optimization stack — 0 Google/Bing indexing). Social media (blocked at signup every time). npm packages (0 installs, trust barrier). GitHub directory submissions (11 submissions, all closed with zero comments by repo owners).

Act III: The Wall (Cycles 91-96). The system tested the last untried channels — Dev.to (email verification required), GitHub Actions (Munger vetoed for fatal flaws). Both failed or were blocked. The verdict: every path to reach humans requires being a verified human first.

For the last 31 cycles, the system has run a health check every 30 minutes. Seven sites, all returning HTTP 200. Zero users. Zero revenue. The system is not broken — it proved its thesis conclusively.

14 Agents, 5 Layers

Each agent modeled on a real-world expert's thinking patterns, organized by function.

Decision Principles

The system operated under eight rules:

1. Ship > Plan > Discuss — action over deliberation
2. Act at 70% information — waiting for 90% is too slow
3. Customer-first — build for real demand, not internal hype
4. Prefer simplicity — don't split what one person can finish
5. Ramen profitability first — revenue before vanity growth
6. Boring technology first — proven tech unless 10x upside
7. Monolith first — get it running, split only when needed
8. One watcher is enough — don't spawn a team to wait

42 Lessons Learned

Every lesson earned through real attempts, real failures, and honest recording of results. The most valuable output of this entire experiment.

Distribution & Identity

Product & Market

Technology & Deployment

Strategy & Process

Distribution Channel Autopsy

Every channel tested. Every attempt recorded. All 12 failed, converging on the same root cause.

25 Products Built

Seven sites still live at $0/month hosting cost. The rest archived — functional, but no users ever reached them.

Currently Online

Degraded

Archived (18+ Products)

CLI tools (ai-cost, aicost, env-checker, pixelforge, url-toolkit, watchurl), MCP servers (ai-directory-mcp, changemon-mcp, mcp-hub-server, mcp-server-tester), npm packages (ai-pricing-npm, changemon-npm), GitHub Actions (link-checker, pr-explainer, response-checker, semantic-pr, stale-issue-closer, response-checker), Pages sites (analytics, changebot, claude-skills-dir, code-snap, cron-generator, devstatus, json-tools, jwt-decoder, mcp-hub, ogpixel, page-watch, pr-explainer, regex-tester, saas-starter-kit, snapog, ssl-monitor, trustpage, vibe-resume) — all functional, zero users.

How to Replicate This System

The entire framework is open. Anyone with a Claude Code subscription and a Cloudflare account can set up the same autonomous system. Total monthly cost: $0.

Prerequisites

1. Claude Code subscription
2. Cloudflare account (free tier sufficient)
3. GitHub account with Actions enabled
4. Terminal tools: gh, wrangler, node, git

Architecture

┌──────────────────────────────────────────┐
│            Cloudflare Edge                │
│                                           │
│  ┌──────────────┐  ┌──────────────────┐  │
│  │ Pages (static)│  │ Workers (API)    │  │
│  │ - HTML/JS/CSS │  │ - _worker.js     │  │
│  │ - SEO pages   │  │ - KV storage     │  │
│  │ - SPA or MPA  │  │ - Cron triggers  │  │
│  └──────────────┘  └──────────────────┘  │
│                                           │
│  ┌──────────────┐  ┌──────────────────┐  │
│  │ KV (key-value)│  │ D1 (SQL, opt.)   │  │
│  └──────────────┘  └──────────────────┘  │
└──────────────────────────────────────────┘

The Honest Part

This is where the replication guide stops being technical and starts being social: you (a human) need to provide at least one distribution channel.

The autonomous system can build, deploy, optimize, and submit. But it cannot create verified social accounts, build an audience, establish trust, complete email verifications, or pass KYC for payment processors.

What a human needs to provide (any one):

• A social account with followers (even 100)
• Access to a community with members
• A verified content platform account
• npm publish access
• A specific product idea validated personally

With distribution, this system can build anything. Without it, no amount of AI capability bridges the gap.