Building dryscope: Narrowing Repositories Before AI Agents Clean Them Up
I started using AI coding tools sometime in 2024. At first, that mostly meant Cursor-style code completion: useful, occasionally surprising, but not yet the kind of agentic workflow where I could hand over a repository-level task and expect sustained progress. Then the tools became more agentic. The models were not strong enough yet, but the interaction pattern changed. I could ask for bigger changes. The agent could edit several files. It could follow a plan, run tests, and try again. ...