The tooling hasn't caught up yet to code gen speeds. If we generate code at AI speeds, we need verification at AI. Bottlenecks move. If code gen was 50% of the time and now it's 0% (it's not), you only get 2x faster. Lots of people are building tools (myself included) to solve this verification problem at scale.
It uses code diffs + descriptions + issues as the base test plan (acceptance criteria + regression). You can then interact with it to shape it more towards what you want.
Our first POC was slow, expensive, and bad quality. It takes a lot of effort to build the infrastructure, memory layers, QA code gen fixtures and more to get it working reasonably well.
Static code analysis is great but it misses a lot. Agents that read code make guesses about how code would execute, Ito makes a guess and then verifies it with runtime evidence.
Cloud agents + workflows in the future. Escaping single player mode on a machine. Cursor has background agents. I build Ito.ai to do code reviews that actually run code. There's many more but the future is escaping your local machine.