Your AI builder burning credits and breaking things as your app grows? Why it happens, what changed in 2026, and how to make your app easier to change.
January. I’d type what I wanted. Wait. Watch the agent work. Then it would stop and ask me a question.
The same question it had asked ten minutes earlier.
I’d answer again. It would carry on, change something it had already changed, and ask again. Every round cost credits. By the end of the build, I’d spent thousands of dollars in credits and weeks of evenings on an app that, on a good day, did what I’d asked the first time.
If you built something in that era, you probably know the feeling.
I blamed myself for a while. I wasn’t asking it the right way. Then I realised the tool itself had limits, and I was paying for them.
Quite a lot, and it’s measurable.
AI can now work on much longer tasks. METR, a research group that tests AI systems, found that the length of tasks AI agents can complete has been doubling about every seven months. (METR)
The builders keep changing, fast. The agent I used was already able to work on its own “for up to 200 minutes”, and to test itself in a real browser. (Replit, September 2025) Running longer didn’t stop it forgetting what we’d agreed. Then, a few weeks after I finished, Replit launched Agent 4, built around planning first: a Plan mode to “map out your build before any code is written”, a Design Canvas to try designs before committing to code, and several tasks running at once. (Replit, March 2026) Plan first, then build is exactly what I was missing in January.
I wanted to see whether the code itself looks different, not just the marketing. So I took 14 public apps built with Lovable, 7 created in early 2025 and 7 created this year, and ran them through the same counting tool Verilay uses. I left out portfolios and practice projects.
| Middle value of 7 apps each | Early 2025 | Mid 2026 |
|---|---|---|
| Copy-pasted blocks per 1,000 lines | 14.4 | 2.4 |
| Code that may be unused | 8.2% | 1.5% |
| Lines in the biggest file | 603 | 1,583 |
The newer apps repeat themselves far less, and leave much less dead code behind. That matches what I’ve felt building with today’s tools.
But look at the last row. The newer apps have bigger single files. Today’s builders write tidier code, and still let files grow.
Fourteen apps is a small sample. Newer apps might be smaller simply because they do less, and plenty of early apps were tidy. So I’m not claiming proof. What I’m comfortable saying is this:
So if your app started in 2024 or early 2025, or it’s grown past 20,000 lines, it’s worth knowing what’s in there.
This is why I’ve added Code fitness to Verilay’s deep scan.
A deep scan reads every file in your repository, not just a sample, and checks it for security problems, layer by layer, with a fix prompt for each one. Now it also answers a second question: is your app getting hard to change?
It doesn’t use jargon. It starts with things you’ve probably noticed:
For each, it shows the number behind it, measured in your code, and the files involved. Then it gives you the three things to tidy first, each with a prompt for your AI builder that starts with “report first, change nothing yet”. One step at a time, so nothing breaks while you tidy.
When you’ve made a change, scan again. The numbers tell you whether it worked, not the builder’s word. (How to fix safely, step by step, is its own series: Fix it with AI.)
If you started building early, especially if your app is past 20,000 lines: does any of this sound familiar? Did changes get slower and more expensive as it grew? Email me at [email protected]. I read every reply.
And if you want to know what’s in your own app, the free scan is at verilay.dev, and the deep scan now includes Code fitness.
One short email a month: what’s new at Verilay, one thing to check in your AI-built app (with a ready-to-paste prompt), and the AI security news worth knowing. No spam, and you can unsubscribe from any issue.