A multi-agent Claude Code "company" ran for thirty days. It shipped nine products and over 280 posts. It earned zero dollars.
The post-mortem is one of the bluntest things I have read from someone running agents at scale:
"9 products built. 280+ posts. Full pipeline. $0 revenue. Multi-agent systems can build anything. They cannot sell anything."
I went looking for the counterexample. I distilled 3,549 evidence clusters out of YouTube, podcasts, GitHub, Reddit, Indie Hackers and dev.to, ranked by how many independent sources corroborated each one. Distribution being the bottleneck, rather than build capacity, is the single most repeated strategic lesson in the entire set. It is not close.
Why this is counterintuitive right now
Building got cheap first. That is the whole story.
If you can produce a working product in a weekend, your intuition tells you the constraint is still production, because production is the part you can see. You watch the diff grow. You feel progress. Selling produces no artifact for days or weeks, so it feels like not-working.
So people build a tenth product instead of finding the first customer for the first one.
The corollary tactic
The most concrete fix in the corpus is a scheduling rule, stated by someone who had clearly lived it:
"when you fire off claw [Claude] to go work for 15 minutes you could do something businessy for those 15 minutes and figure out how the hell you're going to sell this thing you're building"
Time-box selling equal to building. Not "do marketing eventually." The agent is working for fifteen minutes anyway. That is fifteen minutes you are not writing code, and you were going to spend it reading Twitter.
What "selling" means when you hate selling
It does not mean cold calls. In the evidence, the things that actually moved products were unglamorous:
- Can someone pay you today, without talking to you? A live price and a live checkout. This is the one that catches people. A launch that hits the front page with no payment path earns exactly zero, and there are documented cases of precisely that.
- Has a stranger completed the core action end to end? Not you. Not a friend. If nobody outside your own head has finished the loop, you do not have a distribution problem yet, you have an activation problem, and marketing will just buy you a bigger sample of people bouncing.
- How much distribution have you actually sent? With numbers. "I posted about it" and "I sent 40 emails on Tuesday" are different diagnoses with different fixes.
The uncomfortable part
Distribution work does not compound the way building does. A refactor makes tomorrow's building faster. Sending forty emails does not make the next forty easier. It is genuinely worse work, and that is exactly why it stays undone while the codebase gets beautiful.
The nine-products case is useful precisely because the build side was not the problem. The pipeline worked. Nine things shipped. The machine did what it was told and the number at the end was still zero.
If you want the full version of this, I packaged the diagnosis as an MIT-licensed Claude Code plugin: it asks five questions, names one failure mode, and gives you ranked moves with a source URL on every claim. It also ships the graveyard, the eleven popular tactics that died under adversarial re-checking, which is the part I would want if I were reading someone else's list. It is at why-isnt-it-selling.
Top comments (0)