Project progress.
Today, on the coding agent project, my agents completed a 55-ticket and a 40-ticket frontier on the Pizza Empire benchmark, where they’re being asked to design and implement a complete online pizza business, with kitchen tools, delivery tracking, and all the toppings 🍕🍕.
In all, 180,000 lines of code were generated in less than 24 hours.
Best day so far. We can do better.
The new system is successfully swarming across multiple providers simultaneously. This is the joy of Ralphing the workers.
The first thing I had to address was concurrency. I needed dynamic scanning for when workers stop responding because of provider cues. That’s a thing. I feel like the safe number for most providers is somewhere between 5 and 10 threads.
This makes me sad, but I understand why those limits exist.
Still, if I’m paying for API access, I think I should be able to suck down the water hose of data as much as I want, dammit.
Git and non-breaking edits
Today I moved from shadow repos to integration repos.
The git functionality got complicated fast, because swarms will be swarms, and there were a myriad of technical problems between the workers and the git system that I hadn’t considered before this morning.
My commitment to non-breaking edits matters here. It’s a moral pillar.
I could build the project without it, but then I’d be risking accidental code deletion, agents stomping on each other, and the kind of invisible damage that makes long-horizon automation feel broken and slow.
Now, the agents run like little soldiers on their dedicated branches, while the system automatically consolidates finished work and resolves repo conflicts deterministically.
QA’s job is getting more complicated, but this isn’t the kind of QA drift I’ve had in previous systems. This time, the role is still fairly well scoped.
My hope is that this is the last big architectural change I need to make before launch.
But this is why god invented agent evals.
Without them, I wouldn’t have seen any of this until much later, and people would have been disappointed.
I’m all kinds of happy with the state of things.
Testing is running a lot smoother than it was, and GPT-5.5 is showing no signs of scope fatigue at the moment.
I’m feeling like I’m back on track.
More soon. 🕯️