Yeah, that's not the way... It's synthesizing a single hypothesis at a time with attention to the emerging architecture, then decide to reimolement manually or reapply agents to refactor the code until it's reasonable.
The people I've seen do things like this blew through good ideas because the implementations were flawed and they couldn't be bothered to you know read the code because people ... Are told not to do that anymore. In some situations they are not even allowed to read or discuss code at work anymore. It's ridiculous...