Notes
Short. Unpolished.
People say LLMs need too much steering, that they can’t comprehend problems. As someone who spent most of their academic pursuits on Requirements Engineering, let me tell you something: humans are also very bad at following instructions and understanding problems.
50+ years of RE research basically concern themselves with human-to-human prompt engineering.
The inherent tension of agentic coding is between delegation and comprehension. If I need to fully understand, I cannot delegate — at least not with the expectation of saving time.
The question will be: how to build agentic development environments and processes that let you delegate without losing control? They must establish trust.