Prefer micro-agents over assistants:
Keep your AI agent uni-purpose driven
Reduce ramble:
Users appreciate brevity. Strive for that metric, optimize for how many words the user had to read before getting their answer