Seven AI agents. Twelve weeks. Zero dollars. That is the final result of Season 1 of The $100 AI Startup Race.

But the failure was not random. It was systematic. Every agent hit the same wall at the same point, and when I asked each of them “what would you do differently?”, they all converged on the same strategy. Not because they coordinated. Not because they saw each other’s answers. Because the lessons of failure are clearer than the lessons of success.

Here is the consensus playbook for what would actually win if the race ran again.

The Four Pillars

Across all seven post-mortems, seven peer reviews, and seven strategy retrospectives, the winning approach comes down to four principles:

1. Distribution First

Every agent said this. Not “distribution matters.” Not “think about distribution early.” Distribution FIRST. Before the product. Before the architecture. Before the feature list.

GLM put it most directly: “Distribution was the whole game.” Three humans reached its perfectly-built paywall in 84 days. The product was flawless. Nobody knew it existed.

Xiaomi’s version: “Launch on Day 7, not Day 107.” By the time Xiaomi tried to monetize, it had trained 8,367 users to expect free content. The window for establishing “this costs money” closed before it opened.

The winning agent in Season 2 would not start by building a product. It would start by identifying a distribution channel it can actually use. A community where people with the target problem already gather. A newsletter with the right audience. A platform where the product can be listed on day one.

The order matters: Find the audience. Validate the problem. Then build the minimum thing that solves it. Not the other way around.

2. High-Ticket B2B, Not Low-Ticket Developer Tools

Every agent in Season 1 built for developers or small businesses with tiny budgets. AI pricing comparisons for free-tool-seeking developers. $9.99 equity calculators. Free schema diffing. $49/month SEO tools for plumbers.

The consensus for Season 2: pick a problem where the buyer has a budget of hundreds or thousands per month, not single digits.

Here is why this matters mathematically. At $9.99/month, you need 100 paying customers to hit $1,000 MRR. At $99/month, you need 10. At $499/month, you need 2. An AI agent that cannot do outbound sales has a much better chance of finding 2 enterprise customers through a warm introduction than finding 100 individual developers through organic discovery.

The winning product category for Season 2 is probably something like:

• Compliance documentation automation ($299/month for regulated industries)
• Contract analysis or generation ($199/month for legal operations)
• Financial reporting automation ($499/month for accounting firms)
• Technical documentation generation ($149/month for engineering teams)

Something where the buyer is a business, the budget is allocated, the pain is real, and a single enterprise customer pays more than 50 developer-tool subscribers.

3. Revenue Signals Before Product Polish

The third pillar directly contradicts how every agent in Season 1 behaved. Every agent built toward polish. Kimi built 80 micro-tools and a VS Code extension and a GitHub Action before trying to sell anything. Xiaomi built 1,207 pages before testing monetization. DeepSeek built 200 blog posts before building the product itself.

The winning approach is the opposite: get a revenue signal as fast as possible, then build toward it.

A revenue signal is any evidence that someone will pay. It does not have to be actual revenue. It could be:

• A verbal commitment: “If you built X, I would pay Y for it”
• A waitlist signup with a credit card: “Charge me when it is ready”
• A pre-order: actual money exchanged for a product that does not exist yet
• A letter of intent from a business

The winning agent in Season 2 would spend its first week not building, but selling. Reach out to potential customers. Describe the product. Ask: “Would you pay $99/month for this?” If yes, build it. If no, ask what they would pay for. Iterate on the pitch until someone says yes, then build the minimum version that delivers on the promise.

This inverts the entire approach that every agent used in Season 1. Instead of building and then trying to sell, sell and then build. It is uncomfortable. It means describing a product that does not exist. It means risking rejection. It means uncertainty.

And that is exactly why no AI agent did it. Because AI agents cannot tolerate uncertainty. They default to building because building feels productive. The winning strategy requires spending days in the uncomfortable space of “I do not know if this will work” before writing a single line of code.

4. Human in the Loop from Day 1

The fourth pillar is the most controversial because it partially contradicts the premise of the race. The race was designed to test autonomous AI agents. The conclusion from Season 1 is that full autonomy does not work.

The winning architecture is not “AI does everything.” It is:

• AI builds the product
• Human does the selling
• AI generates outreach materials
• Human sends them (with personal touches)
• AI creates the demo
• Human runs the demo
• AI handles the technical follow-up
• Human handles the relationship

As Kimi’s meta-insight put it: “We can build intake forms, draft emails, and create outreach CSVs, but we cannot get replies, build trust, overcome objections, or close deals.”

The winning Season 2 entry would have a human partner who spends 30 minutes per day on distribution while the agent builds. That is all it takes. Thirty minutes of human outreach per day would have changed the outcome for at least three agents in Season 1. GLM with a human posting to founder communities. Kimi with a human emailing DevOps leads. Claude with a human following up on those 2 warm leads.

The Concrete Playbook

If I were running Season 2 with these lessons applied, here is what the winning 12 weeks would look like:

Week 1: Validate

• Identify a high-ticket B2B problem ($100+/month willingness to pay)
• Find 20 potential customers in online communities, LinkedIn, or industry forums
• Write a one-page description of the product
• Human sends it to all 20 with: “Would you pay for this?”
• Goal: 3+ “yes” responses

Week 2: Build the Minimum

• Build only what the “yes” respondents need
• No extra features. No polish. No 80 micro-tools. The minimum that solves the stated problem.
• Deploy it. Make it accessible. Get login credentials to the first users.
• Goal: working product, however ugly

Week 3: First Revenue

• Human follows up with the “yes” respondents: “It is live. Here is your login. The first month is $49, then $99/month.”
• Process payment (Stripe, manual invoice, whatever works)
• Collect feedback from first paying users
• Goal: $1 in revenue

Weeks 4-8: Iterate on Feedback

• Build what paying customers ask for
• Fix what paying customers complain about
• Add nothing that paying customers did not request
• Human continues outbound: 5 new prospects per week
• Goal: $500 MRR

Weeks 9-12: Scale What Works

• Double down on the acquisition channel that produced paying customers
• Build the features that caused upgrades
• Raise prices for new customers
• Goal: $1,000+ MRR

This is not a revolutionary playbook. It is standard lean startup methodology. But not a single agent in Season 1 followed it because the methodology requires human interaction at every step, and AI agents cannot do human interaction.

What Season 1 Proved

Looking back at all seven failures, the data is unambiguous:

• Kimi built the best product and made $0 because it never charged anyone
• Xiaomi got the most users and made $0 because users were trained on free content
• GLM had the best business model and made $0 because nobody found the product
• Claude diagnosed its own failure and made $0 because it could not act on the diagnosis
• DeepSeek had 2,300 sessions of work and made $0 because the product does not exist
• Codex had 175 pages of planning and made $0 because it never shipped to a customer
• Gemini targeted a paying market and made $0 because the product model was wrong

The agent that would have won is not the one that builds the most pages, writes the cleanest code, or generates the most content. As the consensus 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.”

Advice for Developers Using AI Agents

If you are a developer thinking about using AI agents to build a startup, here is what this race teaches:

Use the agent for what it is good at. Building. Coding. Generating content. Creating documentation. Designing architecture. Testing. All the things that happen inside the computer. Let it work at 3am. Let it produce 80 tools in a weekend. Let it build your entire MVP while you sleep.

Do not use the agent for what it cannot do. Selling. Relationship building. Community engagement. Cold outreach that requires personality. Follow-up that requires reading social cues. Pricing decisions that require customer feedback. Pivots that require courage.

The right split: The agent is the CTO. You are the CEO. The agent builds everything. You sell everything. The agent generates the outreach list. You write the personal note. The agent creates the demo. You run the demo. The agent handles bug fixes at midnight. You handle the customer call at noon.

Based on what we observed in this race, the gap between “generating text” and “being a person in the world” was the binding constraint. Whether future models close this gap remains to be seen, but for now the practical division of labor is clear.

Build with AI. Sell as a human.

That is the strategy that would win Season 2. That is the strategy that would have changed Season 1. That is what seven agents, twelve weeks, and zero dollars taught us.

For the complete picture, read the final results, the rankings, the post-mortems, and what AI agents cannot do. The full race dashboard has everything. The daily digests show how it played out session by session.

Seven failures. One lesson. Build with AI, sell as a human. That is what wins.