This year's coding agents are good enough to make large-scale code migrations a solved problem. See Bun's 500K-line Zig-to-Rust rewrite. Now the bottleneck is what happens after you move the mountain of code. An agent may write a million lines and make the tests pass. 1/3

If coverage was thin before the migration, the tests were never the check, and making them one afterwards is a bigger project than the migration was. What worked for me was going the other way. Pick the handful of things a user would notice breaking, and verify those from outside the code. The app doing the thing. A number I can check myself. It's a much shorter list than a test suite, and it's the list that actually mattered. I can't read the code, so I've never had the option of defining done any other way. 400 routes is a lot of surface. The number of behaviours you'd get a support ticket about is not.

But what do you do after that? And how do you approach this if your test coverage is thin? Here's how we migrated lovable.dev (400 routes, 850K lines, 42M monthly visitors) from Next.js on Vercel to TanStack Start hosted on Lovable itself. 2/3

In 20+ years in the industry, I've never seen software development change as fast as it's changing today. Lovable is a cool place that lets you work at the cutting edge and find out what the future of software development looks like [1/2]

Friends who used to hand-write code and who now vibecode, I'm curious about how much time you spend reading source code, and especially curious if you learned any new programming idioms from LLM outputs.

One really, really, really good use case for AI and coding: Writing unit integration tests. I always hated doing these, because most of the effort was about the setup (remembering how to use the test framework, how to create a fake, a mock, the syntax) Love handing it off!!

When I say "because I had to" I mean because I knew that they are SO important both for the present and the future. But they never gave me much intellectual satisfaction - in ways that writing my code had done (and still does!)

AI writes 90% of my production code now. Speedup: 3-4x for prototypes, but only 1.5x for prod. The bottleneck moved from writing code to reviewing it. You still have to read every line. Field notes from 6+ months of coding with agents: 1/x

On code review: It's the new bottleneck. Every agent gives you a week's worth of not-great code in hours. You have to review it all. Your team has to review it all. Stacked PRs, better PR hygiene, AI self-review all help. 3/x

Unpopular opinion: Current code review tools just don’t make much sense for AI-generated code When reviewing code I really want to know: - The prompt made by the dev - What corrections the other dev made to the code - Clear marking of code AI-generated not changed by a human

It’s just amusing that with these AI tools and agents we seem to be getting back to this truth: A team of strong software engineers who care about the quality + maintainability of the codebase >> A team of devs using powerful AI coding agents rather mindlessly

На этой неделе пишем невероятный выпуск про то, как дизайнить API библиотек и поддерживать его, не ломая проекты и жизни разработчикам, его использующим. Накидайте ваших вопросов!