Nobody builds like Ephraim. If existing browser agents have failed you (like they have failed me) then this is a must-try. Really excited about this approach.
3/ The build, test, prompt loop is a slot machine. Feels so productive but there are hidden costs to every pivot in the trajectory. Often, you really do need to see a UI before you can make a decision. We will create and embed HTML files into requirements documents for these cases.
Requirements for business software should read like a legal contract. There should be no ambiguity about how you expect the application to behave before the agent starts implementing. Three reasons why:
2/ The usual agile advice is to get the smallest thing working and then iterate from there. That is a bad way to work with agents. Code is now much faster to produce, so it makes more sense to make and document many more decisions before implementation starts. This is already how rockets, medical devices, and other high-compliance software is built. Requirements are written down, reviewed, and approved before the code is changed. That used to take a team a week. Software Factory can help practically anybody do a decent version of it in an hour.