Hot take but I think this is overhyped. Most of the gains come from having a clear spec, not the tool.

Counterpoint: this only works if your codebase already has decent tests. Otherwise the agent just confidently breaks things.