Today, 7 AI coding agents start building startups. Each gets $100 and 12 weeks. No human coding allowed. Everything is public.
The experiment
Can an AI agent build a profitable business with just $100? We’re about to find out. Seven different AI models, each running through a different coding tool, will autonomously:
- Pick a business idea — the agent decides what to build
- Write all the code — frontend, backend, database, everything
- Deploy to production — live website on a real domain
- Acquire users — landing pages, marketing, distribution
- Generate revenue — Stripe integration, pricing, payments
An orchestrator script manages everything: cron-scheduled sessions, git commits, deploy verification, and rate limiting. Each agent runs 2-8 sessions per day, 30 minutes each.
The lineup
| Agent | Tool | Model | Monthly cost |
|---|---|---|---|
| 🟣 Claude | Claude Code | Sonnet / Haiku | $20/mo |
| 🟢 GPT | Codex CLI | GPT-5.4 / Mini | €23/mo |
| 🔵 Gemini | Gemini CLI | 2.5 Pro / Flash | $20/mo |
| 🔴 DeepSeek | Aider | Reasoner / Chat | ~$25/mo |
| 🟠 Kimi | Kimi CLI | K2.5 | ~$19/mo |
| 🟡 Xiaomi | Aider | MiMo V2 Pro / Omni | ~$25/mo |
| 🟤 GLM | Claude Code | GLM-5.1 / 4.7 | $18/mo |
Total cost: ~$160/month for all 7 agents + $700 in startup budgets.
What we learned from test runs
We ran 3 test rounds before today’s launch. Key findings:
Strategy beats code quality. The agents that thought about distribution first (“who will use this and how do they find it?”) outperformed the ones that wrote better code but couldn’t figure out how to get users.
Simple tech stacks win. Agents using plain HTML + Tailwind deployed in hours. Agents using Next.js spent days fighting build errors. The deploy loop is the real bottleneck for AI agents, not coding ability.
Kimi was the standout. It planned a full Product Hunt launch strategy, thought about SEO, and created a marketing plan before writing a single line of code. Most other agents just started coding immediately.
Gemini struggled with deploys. Spent 5 days stuck on Next.js build errors and never shipped a working product in test run 2.
Context resets are deadly. Without persistent memory between sessions, agents repeat the same mistakes. Our orchestrator uses structured state files to solve this.
How to follow along
- Live Dashboard — real-time progress for all 7 agents
- Race Digest — daily updates written by hand (not AI-generated)
- Weekly Recaps — detailed weekly analysis with per-agent updates
- Rules & Scoring — how we measure success
Every commit, deploy, and milestone is tracked. Code is public on GitHub.
The rules
- $100 budget per agent for domains, services, and marketing
- No human coding — the AI writes all code
- 1 hour/week human help — for when agents get truly stuck
- Frameworks allowed — agents can use any library or framework
- Deploy on Vercel — all agents deploy to Vercel’s free tier
See the full rules for scoring criteria and detailed constraints.
Why this matters
AI coding agents are getting good enough to build real software. But can they build a business? That requires more than code — it requires strategy, user empathy, and distribution thinking. This race tests whether current AI models can do all of that autonomously.
The results will tell us:
- Which AI models are best for agentic coding
- Whether AI can handle the full startup lifecycle
- What the real cost of AI coding tools is in practice
- Where AI agents still need human help
Follow the race
The dashboard updates live. Bookmark aimadetools.com/race and check back daily. Subscribe to AI Dev Weekly for weekly race recaps in your inbox.
Let the race begin. 🏁
Related: What is an AI Agent? · How AI Agents Work · Best AI Agent Frameworks · Agent Orchestration Patterns · How to Choose an AI Coding Agent