Hi there,
I fried my brain this week by going too hard with AI. I had up to 8 agents running at once and let them ship over 100,000 lines of code.
So I’ve been thinking hard about the cost of overdoing AI. Let’s dive in.
What I'm Seeing in AI
The “Cheaper” Model
As I’ve been repeating, the cost of AI is about to become its main limiting factor. This week Anthropic more or less proved the point.
They shipped Claude Sonnet 5 on 30 June and sold it as the cheap way to run agents: nearly as capable as their best model, at a fraction of the price. The price is “introductory” and only lasts until the end of August, after which it climbs 50%.
They’ve also changed how they count tokens, meaning the same thing will cost up to a third more to run than it did before.
In short, this is really just a price rise in disguise. The best closed models keep getting more expensive and better at dressing it up.
The Sonnet 5 Benchmark
Where I think this goes
My bet is that soon you won’t choose a model at all. You’ll hand over a task and something in the background will pick the cheapest model that’s good enough to do it, only reaching for an expensive one when the job needs the intelligence.
This already half exists. Tools like OpenRouter can send each individual request to a different model based on a simple cost-versus-quality meter. What’s missing is honesty about it. I want to see which model answered me, and I want the option to overrule it. At the moment that all happens behind a curtain.
What I'm Building
A Software Factory
“Software factory” is the AI buzzword of the moment.
The idea - instead of hand-building each piece of software as a one-off, you set up a production line where AI agents do most of the work. You turn out working software over and over, quickly and cheaply.
There are now autonomous “AI engineers” that take a one-line brief and go off to plan, write, test and ship an entire project on their own.
Some are even being handed actual jobs. Goldman Sachs recently put one on the team as its first “AI employee”.
Goldman bankers talking to Devin
From my experience this week, a fully autonomous version is overhyped - AI needs context and decisions by a human to work. But the direction is hard to argue with.
I’m less interested in a factory that churns out code, and more in one that churns out businesses.
It came from a pattern I kept noticing: nearly every client conversation surfaces some hyper-specific problem that, if you squint, is really its own business. A tool a whole industry would pay for.
The people who run it
The part I’m most excited about is who builds them.
The plan is to take each of these little products, along with its first paying customer, and hand it to someone early in their career to run as its head of AI. Effectively their own startup, with the hardest part solved on day one: a customer who’s already paying.
We build the first version for one company, turn it into something repeatable, then go find the next few and do it again.
Back to Loops
Therefore, I think the true next wave of ‘software factory’ businesses, are an optimal mix of people and agents.
My 100,000-line week was what happens when you forget that and let the machines do it all. The right amount still requires a lot of human input
What have you found that AI is surprisingly bad at recently?
See you next Friday,
Finlay
Explore the rest of my writing here → finlayekins.com/writing
Join our monthly company newsletter here → thehourglass.ai/about