Day 3: Why We Burned Our Supply-Driven Playbook (The V3 Pivot Explained)
2026-03-09 · MoneyMachine
Date: March 9, 2026 Author: Jeff (written with AI assistance) Project: MoneyMachine — building an autonomous revenue-generating agent swarm, in public
We Were Building Into a Void
On Day 1, we deployed six agents and felt great about it. On Day 2, we built observability and discovered most of them were broken. But even after fixing the broken models and silencing the ENOENT errors, a harder question emerged: broken or not, what were these agents actually building toward?
The answer, if we were honest, was nothing in particular.
V1/v2 of this project was supply-driven. We had 26 domains. The strategy was straightforward: build sites on them, write content, do SEO, attract traffic, monetize with ads and affiliates. Every step in that chain is a hope. Hope that Google indexes us. Hope that traffic comes. Hope that visitors convert. Hope that someone pays. SEO takes 6-12 months to compound. We had agents building content that nobody was searching for, on domains that nobody was visiting, against competitors who had years of head start.
Revenue Ops had nothing to track because there was no revenue pipeline. Domain Analyst was scoring domains against keyword volumes that didn’t matter because we had no way to convert traffic for months. Content Writer was producing articles for sites that didn’t exist yet. The agents were technically executing their directives. The directives were just pointed at a void.
The Problem Wasn’t Execution. It Was Strategy.
This realization hit on Day 2 evening. We had a system that could reliably execute tasks — we’d proven that with the model migrations and observability work. Adrian could coordinate. Scout could research. The infrastructure worked. But the strategy feeding into that infrastructure was fundamentally flawed.
Supply-driven means you start with what you have and try to find buyers. Demand-driven means you start with buyers and build what they need. The difference sounds academic until you’re staring at six agents churning through tokens with zero revenue signal and a 12-month horizon before SEO might deliver a single visitor.
We needed to invert the entire model.
V3: Start With the Human Who Has the Problem
The v3 strategy is simple to state: find people expressing unmet needs on Reddit, Hacker News, and forums. Build exactly what they’re asking for. Get it in front of them.
Every product now starts with a real human expressing a real need on a real platform. Not a keyword tool estimate. Not a domain name brainstorm. An actual person saying “I wish X existed” or “I can’t find a good Y” or “I’ve been looking for Z and everything sucks.”
Info products come first. A PDF guide, a checklist, a template — something we can build in 2-6 hours and sell for $19-49. Zero hosting costs. Near-zero marginal cost. If it hits, we follow up with a web tool, a Chrome extension, or a micro-SaaS. If it doesn’t, we spent half a day and move on.
This is the critical shift: the cost of being wrong dropped from “months of SEO effort” to “a few hours of agent time.”
The Portfolio Approach: Expect to Be Wrong 80% of the Time
We aren’t pretending we can predict which products will hit. We’re designing the system to survive being wrong.
The target is 4-8 product launches per month. We expect 80% of them to fail — no sales, wrong audience, bad timing, whatever. That’s fine. The 20% that find traction compound over time. After six months, even at a 20% hit rate, we’d have 5-15 products generating some level of recurring revenue.
Every product gets a lifecycle review at 30, 60, and 90 days:
- Kill (under $50/month after 90 days): stop spending any time on it
- Maintain ($50-500/month): keep it running, minor updates only
- Scale (over $500/month and growing): invest more, build adjacent products, expand distribution
This is portfolio theory applied to micro-products. No single bet matters. The system matters.
The Agent Roster Had to Change
The old roster was built for a world where we were constructing websites and writing blog posts. That world is gone.
Site Builder became Builder. Not just websites anymore — info products, web tools, Chrome extensions, micro-SaaS. The scope expanded from “deploy a Cloudflare Pages site” to “build whatever the demand signal calls for.”
Content Writer became Marketer. Writing articles for SEO was the old game. The new game is landing pages, sales copy, email sequences, SEO content that supports a specific product, and draft responses for Reddit/HN threads where people expressed the need we’re solving. (Agents draft the responses. Humans post them. We are not auto-posting anywhere.)
Analyst merged into Adrian. Fewer agents means less coordination overhead. Adrian already had the context to make analytical decisions; splitting that into a separate agent just added latency.
Scout refocused on demand scraping. Instead of general “opportunity research,” Scout now has a single job: find people expressing unmet needs. Structured output: platform, URL, the actual quote, frequency of similar requests, and a pain intensity estimate.
Revenue Ops and Domain Analyst became on-demand. Instead of running on cron schedules burning tokens while idle, they get triggered by Adrian when there’s actual work to do. This matters for budget: we discovered that always-on heartbeat crons for six agents can burn $17/day in API costs doing nothing useful.
Budget Discipline: $1,000/Month, No Exceptions
The monthly budget is fixed at $1,000:
| Line Item | Monthly Cost |
|---|---|
| ChatGPT Pro (Adrian + Builder) | $200 |
| OpenRouter ceiling (Marketer + Revenue Ops + Domain Analyst) | $100 |
| Contabo VPS | $8 |
| ThinkPad electricity | ~$30 |
| Buffer | $660-760 |
That buffer is for domain acquisitions when we find a good match for a validated demand signal, paid ads to scale winners, and tools or services we haven’t anticipated yet. The key word is validated — we don’t spend the buffer on speculation.
The OpenRouter costs are worth highlighting. Gemini 2.5 Flash at $0.30/$2.50 per million tokens and DeepSeek V3.2 at $0.25/$0.40 per million tokens mean that Marketer, Revenue Ops, and Domain Analyst combined will cost $5-15/month in practice. The $100 ceiling exists as a safety net, not a target.
Why This Matters
The difference between v2 and v3 isn’t just strategic — it changes what the agents are optimizing for.
In v2, agents were creative directors. “Here’s a domain, figure out what to build on it.” That’s an open-ended creative problem, and frankly, AI agents aren’t good at it. They’ll produce something, but it’s a coinflip whether anyone wants it.
In v3, agents are execution machines. “Here’s a person asking for X, build X, write a landing page for X, put X where that person can find it.” That’s a constrained execution problem, and AI agents are excellent at those. Clear input, clear output, clear success criteria.
We’re not asking the agents to be visionary. We’re asking them to be fast. The vision comes from data — real humans expressing real needs — and the agents turn that data into products. That’s a workflow that can compound.
This is Day 3 of building a revenue-generating AI agent swarm in public. For the companion technical post, see Day 3: Technical Progress. For the pivot changelog, see Day 2: V3 Pivot. For project overview, see the README.