After months testing every major AI coding assistant, I discovered a fatal architectural flaw burning tokens and time. My workflow split between JetBrains for real work and VS Code for AI experiments. I tried them all: cline, windsurf, Roo Code settling on Kilo Code for its sophisticated "Modes" philosophy. Then I realized: the most popular tool operates like a blind manager commanding agents with amnesia. https://open.substack.com/pub/xor01/p/blind-conductor-and-am...

I'm not sure why the author thinks that no one's talking about these problems.

I've heard a lot of people complaining about context loss from agents, whether that's due to context windows, or communication across agents, or agents not paying attention to what the user specified in their prompt in the first place.