Korva is a scroll of the internet’s best ideas.
The whole company flew to Brazil
Where agents could work
After the first demo went viral, pilots started, and they mostly failed. The agents ran on GPT-4 and were primitive, fine for toy projects and not ready for real codebases. So the team asked which task an agent could do end to end first. Mundane, repetitive work was too simple to need an agent, and deep architecture problems were too hard. In between sat migrations and version upgrades, like moving a company from Java 7 to Java 8: a 50,000-file codebase where the same eight things change in every file, with some thoughtfulness about the trade-offs.
The trip to Brazil
Nubank was the first success, and the use case was one of those big migrations. Scott Wu and the whole team flew to Brazil and sat beside the bank's engineers, learning what to do in each case, what Devin needed to read, and how to debug their exact problems. They tried tasks by hand first to see what each looked like, then taught Devin to do them correctly and built the orchestration around it.
Enterprise customers
Building the product was almost like building it for one company. Fortune 500 firms became the natural market from there, because Walmart, CVS, JP Morgan, and Mercedes all run massive teams of software engineers.
Why are code migrations good work for agents?
A migration repeats the same kind of change across thousands of files, which suits a tireless worker. Success is also easy to check, since the code either compiles and passes its tests or it does not. That gives an agent a clear signal to retry against. Tasks that are open-ended or judged on taste lack that signal.
What changed between Java 7 and Java 8?
Java 8, released in March 2014, added lambda expressions and the Streams API. Those features changed how developers write loops and handle collections. Old code kept working in most cases, but teams moved to the new style gradually. A large codebase could have thousands of places where the older patterns appeared.
Why do big banks carry such large codebases?
Banks have run software for decades, and each new product, regulation, and acquisition has added code on top of what already existed. Removing old code is risky when it handles money, so it tends to stay. Nubank, founded in Sao Paulo in 2013, is among the newer banks, though its codebase has still grown to be very large. Large institutions often keep a team of engineers just to maintain and upgrade the languages and frameworks they already use.
More from David Senra
