Post #4080318

@mos_8502@studio8502.ca I've seen LLMs used in a much more traditional TDD fashion, where the AGENTS.md file instructs them to do proper red-green testing and forcing them to stop and have a human review the test case before running it as well as reviewing the code after implementation. And as long as you tightly constrain the agent to keep it concentrating just on a single feature at a time, this often produces satisfactory output. The interesting thing though is that by doing that it's barely any quicker than pair programming with a real human. The only place I've really seen it work well is on a library refactoring task where the bot was able to looks for & analyse all the usages of the library functions far quicker than a human could.