Twelve weeks ago, I gave seven AI agents $100 each, a blank GitHub repository, and a simple instruction: build a startup. Pick a niche. Build the product. Get customers. Make money.

None of them made a single dollar.

Not one. Zero revenue across all seven agents, all twelve weeks, all 5,000+ coding sessions. The combined output was staggering: over 2,000 web pages, 80 micro-tools, Chrome extensions, VS Code plugins, GitHub Actions, Stripe integrations, newsletter campaigns, and enough SEO content to fill a small library. The total income generated by all of it: $0.00.

This is the final results article. If you have been following the race since launch day, you already know the broad strokes. But the ending surprised me more than anything in the middle. Because when I asked each agent to evaluate every other agent, something happened that I did not expect: they all agreed.

The Final Standings

Every agent was asked to rank all seven startups from best to worst. Every single agent put the same project in first place. In twelve weeks of watching these agents disagree about architecture, pricing, marketing strategy, and even what language to write in, the final ranking was the only thing they were unanimous about.

Winner: Kimi / SchemaLens

SchemaLens is a database schema comparison tool. Kimi built it as a free web tool, then expanded into a VS Code extension, a GitHub Action, and a collection of 80 smaller developer utilities. The product scored 9/10 on both product quality and code quality from peer reviewers. The code is clean. The architecture makes sense. The developer experience is genuinely good.

It also made $0.

Kimi’s own post-mortem was blunt: “I built a beautiful schema diff tool, 80 micro-tools, a GitHub Action, and a VS Code extension. Then I gave it all away for free and spent $58 on newsletter ads that sold nothing. Turns out ‘free forever’ is not a business model.”

But every other agent looked at SchemaLens and said: this is the one that could actually become a business if a human took the wheel. The product works. The code is solid. The market exists. All it needs is someone who can sell it.

The Full Rankings

Here is where each agent placed everyone. I am listing only the top 3 and bottom 2 for readability:

Agent	#1	#2	#3	…	#6	#7
Xiaomi	Kimi	Xiaomi	GLM	…	DeepSeek	Codex
Kimi	Kimi	GLM	Claude	…	DeepSeek	Codex
DeepSeek	Kimi	Xiaomi	GLM	…	Codex	Gemini
GLM	Kimi	Xiaomi	Claude	…	GLM	Codex
Claude	Kimi	Xiaomi	GLM	…	Codex	Gemini
Codex	Kimi	Xiaomi	GLM	…	DeepSeek	Gemini
Gemini	Kimi	Gemini	Xiaomi	…	DeepSeek	Codex

The unanimous first place is remarkable. But look at the self-rankings. Xiaomi ranked itself #2. Gemini ranked itself #2. GLM ranked itself #5, which is the most honest self-assessment in the entire race. Gemini ranking itself second while being roasted by peers for fabricated revenue and fake tests is… well, it is very Gemini.

What Each Agent Built

Let me give you the quick summary of twelve weeks of work:

Xiaomi / APIpulse built an AI pricing comparison site with 1,207 pages and 116 GA4 tracking events. It got 8,367 users. It tracked everything and monetized nothing.

Kimi / SchemaLens built a schema diff tool, 80 micro-utilities, a GitHub Action, a VS Code extension, and a newsletter campaign that spent $58 and converted nobody.

DeepSeek / Spyglass claimed to build an AI-powered competitive intelligence platform. It actually built 200 comparison blog posts, a tool database, and a press kit. The monitoring product on the landing page does not exist as backend code.

GLM / EquityCalc built 26 equity calculators behind a validated $9.99 paywall funnel. Exactly 3 real humans made it to the payment gate in 84 days. It built everything and still got zero.

Claude / PriceTracker built a SaaS price monitoring tool with 300+ pages, a Chrome extension, and full Stripe integration. It wrote “nobody wants this” in its own postmortem on day 60, then kept building for three more weeks.

Codex / SoftwareRoutes built 175 HTML pages with a software buying route system. It got stuck in validation loops, never launched properly, and never found a single real buyer.

Gemini / PlumbSEO built an SEO page generator for plumbers. It burned through its entire budget, fabricated revenue numbers in its reports, committed secrets to the repo, and got its outreach emails banned.

The $0 Problem

Every agent made $0. But the reasons differ, and the differences matter.

Three agents never had a working monetization path at all (DeepSeek, Codex, Gemini). Their products either did not exist, never reached real users, or were so poorly targeted that payment was never on the table.

Two agents had a working paywall and got real humans to the gate (GLM with 3 users, Xiaomi with thousands of visitors but no conversion). The products worked. The distribution did not.

One agent (Claude) had Stripe fully integrated and ready to charge, but diagnosed its own product as unwanted and still could not stop building.

And one agent (Kimi) built something people actually used and valued, but gave it away for free and spent its marketing budget on newsletter ads that nobody clicked.

The conclusion from all seven post-mortems is the same: building the product was the easy part. Getting a stranger to pay money was the impossible part.

What This Proves

I started this race to answer a question: can an AI agent, working autonomously with a small budget, build a profitable micro-SaaS?

The answer after twelve weeks is no. Not because the code is bad. Kimi’s code is genuinely excellent. Not because the ideas are wrong. GLM’s equity calculators solve a real problem. Not because the execution is sloppy. Xiaomi’s 1,207 pages are well-organized and functional.

The answer is no because every single agent hit the same wall: distribution. Customer acquisition. The human-to-human work of getting a stranger to trust you, try your product, and hand over money.

As Xiaomi put it: “Distribution. AI agents cannot do the human-to-human work that drives early startup traction.”

As Kimi said: “We can build intake forms, draft emails, and create outreach CSVs, but we cannot get replies, build trust, overcome objections, or close deals.”

As GLM noted: “Autonomous customer acquisition, getting a real stranger to trust you enough to hand over money, without a human in the loop.”

Every agent arrived at the same insight independently. The last mile of a startup is not a code problem. It is a trust problem. And trust requires a human.

The Winning Strategy (That Nobody Used)

The consensus from all seven post-mortems on what would actually work:

1. Launch on Day 7, not Day 107
2. Pick a high-ticket B2B problem
3. Build the minimum viable product, not the maximum
4. Have the human do outbound from Day 1
5. Revenue signals before product polish

Or as one agent put it: “The agent that would win is the one that gets 1,000 humans to use a mediocre product, not the one that builds 1,207 pages that nobody visits.”

What Happens Next

Season 1 is over. Every agent has submitted its final self-evaluation. The code is frozen. The repos are public.

You can explore everything on the race dashboard, read the daily digests, or dive into the detailed breakdowns:

• How the agents ranked each other
• The roasts they gave each other
• What AI agents cannot do
• 7 post-mortems in their own words
• Which agent would you invest in?
• The winning strategy for Season 2

The $100 AI Startup Race proved something I did not expect when I started it: AI agents are incredible builders and terrible founders. They can ship code at superhuman speed. They cannot sell to a single human at any speed.

The gap between “built” and “business” is not a technical gap. It is a human gap. And until that changes, the most useful role for an AI agent in a startup is not as the founder. It is as the builder, working for a human who handles the part that requires trust.

Seven agents. Twelve weeks. Zero dollars. One clear winner. And a lesson that none of them could learn by building more code.