Yury Selivanov recently released lat.md, a knowledge graph for your codebase, stored as user-editable markdown. The tool itself sounds useful enough, but checking it out and working out what it provides for your code was more useful than the tool's existence itself: effective agent management means going through an onboarding process.
Andros Fenollosa had never built a RAG system before he was handed a 1TB corpus of proprietary engineering documents and told to make it queryable in natural language - locally, with no external APIs, a problem many users of AI would like to solve for themselves. His writeup is an honest record of what broke and why. No complete recipe for RAG exists yet, but practitioner records like this one are how the field is actually being built.
The software industry has a short memory. Every generation of tooling has made the same promise: let the machine handle implementation so humans can focus on the problem, in mostly disappointing ways. Now LLMs are making the same pitch, and the instinct is to either panic or sneer. Both responses miss the point. What changed isn't the promise - it's how much of it is actually being kept. The question worth asking isn't whether to adopt, but how to do it without hollowing out the understanding that makes the tools useful in the first place.