An agent opened a pull request in our codebase. It compiled, the tests passed, it did what the task asked, and it replaced one hook with a raw fetch. That hook was handling deduping, auth, and cache invalidation on publish. None of that was written down. An engineer caught it because he'd seen it go wrong before.
This talk is about what we did next, which was getting that knowledge out of people's heads and into files the agents read. I'll show what a rule looks like, how each one names the command that proves it and the date we delete it, and what a set of them costs to keep honest.
We were building Source at the same time, where the code is the real artifact our users own instead of an export of one, and that's a transformation happening across industries.
You'll leave with the format we use and a way to find your first rule: ask two engineers the same question and see whether they answer it the same way.
Sobre Utkarsh Sengar: Utkarsh Sengar is Chief Technology Officer at Webflow, working on one of the stickier, more interesting problems in software today: what design and code become once AI agents are first-class builders alongside humans. He's leading how Webflow builds for an agent-native web -nothing walled off, code that's real and portable, not locked into a closed system, and agents that can see and act on a project with the same fidelity a human designer has on the canvas. It's less "add AI features," more architecting how software gets built when your co-builder isn't human. He joined Webflow in 2022, following senior engineering leadership roles at Upwork, OpenTable, and eBay. He also advises early-stage founders through seed funds, and holds a Master's in Software Engineering from San José State. Outside of Webflow, he's usually hacking on side projects, or running with his wife and six-year-old. Ver menos