Author here. Uncle Bob recently said he no longer reads his agents' code and relies on a "gauntlet" of tests instead. I disagree: you should read all of it, which means getting agents to write less code that's already been checked against what you meant. The flow, roughly as Microsoft's NFV does it: 1) capture the intent of a change, 2) freeze it as properties, 3) an agent writes the code and a proof, 4) Lean or Dafny checks the proof, 5) you read the properties, then the code. Most of the post is about what such a proof still can't tell you.
saviorand•4m ago