Building Forge: what we learned shipping an AI-native code editor
Forge went from a rough internal prototype to a full AI app-building experience in a matter of months. Our engineering lead shares the architectural bets that paid off — and the ones that didn't.
When we started building Forge, the hardest design question wasn't "can an AI write code" — by 2025, that part was largely solved. The hard question was: how do you make an AI coding agent feel trustworthy enough that someone will let it touch a real project?
Our first instinct was to make every change reversible, and that turned out to be the single most important decision we made. Forge checkpoints state automatically before any meaningful change, which means users can experiment freely with Agent mode without the fear that's kept a lot of people from trying AI coding tools in the first place.
The second bet was on transparency over magic. Rather than hiding what the AI was doing behind a black box, we built three distinct modes — Ask, Agent, and Plan — so users could choose how much autonomy to hand over, and always see the reasoning behind proposed changes before they landed.
Not every bet paid off immediately. Our first version of the visual database builder generated technically correct but awkward schemas — it took several iterations of watching real users struggle before we found a model that matched how people actually think about their data. That's probably the biggest lesson: ship the part you're confident about, and watch closely for the part you're not.