One Person, One Product, One Purpose

The last time I shipped a product this size, it went like this. Months building solo until I hit the wall. Then the offshore hire, the onboarding, the sync calls, the code that kept drifting from what was in my head, and a product that finally landed in Q3 of a year that had started with so much momentum. I lived that version at Intridea scale, and again through years of shipping with hired teams.

This time it took weeks. Alone. And I’m not running one product, I’m running several at once.

Some history, because this shift is aimed straight at people like me. I’ve always been an ideas person. The part I love is the crude prototype, the couple of weeks that prove the thing is actually feasible. Call it 80 percent of the invention and 20 percent of the work. The rest, the polish, the edge cases, the billing, the boring infrastructure that turns a demo into a product, always took the longest and cost the most, and it was the part I had to hire for. Every idea carried a hidden price tag: a team, a runway, a calendar. Knowing how the thing should be built didn’t lower the price. Knowing the market didn’t either. So the ideas queued, and realistically I pursued one at a time, because money and time said so.

AI inverted that. The last 20 percent, the stretch I used to dread and delegate, is now the part that moves fastest, and finishing became fun instead of a hiring problem. The prototype is still the joy. It just no longer ends where the payroll used to begin.

The interesting part isn’t that AI made me a better programmer. It didn’t, particularly. The interesting part is what disappeared: the coordination tax.

Communication paths grow as n(n−1)/2. Two people share one channel. Ten people share forty-five. Brooks said the consequence out loud fifty years ago in The Mythical Man-Month: adding people to a late software project makes it later, because the people who were producing stop producing to onboard the people who aren’t yet. You never bought output when you hired. You bought output minus a tax that grows faster than headcount. Conway’s Law is the same disease showing up in the product: you ship your org chart. One mind produces one coherent architecture with one taste running end to end. Six minds produce a negotiated settlement, and you can feel the seams.

For seventy years we paid the tax anyway, for a good reason. One person couldn’t produce enough. Parallelism was the only way to scale output, and the tax was its price. What changed is that one person’s ceiling went up far enough that, for a whole class of work, the parallelism is optional. And if you don’t need the parallelism, you don’t pay the tax.

But the speed framing undersells it, and this is the part that gets lost. I didn’t get five times faster at one thing. I started building five things. A team is a concentration of force: correct when the target is known and the job is executing one hard, integrated build. Solo plus AI is a distribution of force: correct when the target is unknown and the job is discovery. Most early-stage work is discovery. Five cheap shots beat one expensive bet when you can’t see the goal yet.

Now the honest edges, because a take that hides where it breaks isn’t a take.

This works from zero to one and strains after. Support, reliability, security, the infinite maintenance tail: that’s exactly where teams earn their cost back. One person is also one point of failure, and a team is partly insurance. If the thing needs to outlive my attention, that insurance is worth paying for. I’ve written about when I’d still hire, and I mean it.

There’s also an obvious tension in the title. How can five products be “one purpose”? It resolves more cleanly than it looks. Each product is one coherent stream in one head, and I time-slice between them. What got removed is the inter-human tax, not context-switching. Switching my own contexts is cheap. Synchronizing six people’s version of one context never was.

And running multiple agents does re-import a piece of the problem, the architectural piece. I still have to decompose the work into components that don’t collide, and I’m still the integration layer. What it doesn’t re-import is the social piece. There are no meetings about the boundaries. I just decide.

So the bottleneck moved. Building is cheap now, which means the constraint isn’t “can I afford and coordinate a team.” It’s “can I choose well.” What’s worth building, which of five working versions to keep, when done is done. The model can’t hand you that part. Whether my judgment is good enough to deserve this much cheap execution is the question I get to live with now, and I honestly prefer it to the old one.