Progressive Skill Loading: 40+ Agent Skills, No Bloat

On June 23, 2026, a developer's side project hit #1 on Hacker News: a Google Workspace CLI that exposed 40+ structured \"skills\" an AI agent could call to drive Gmail, Drive, and Calendar. He was fired within days, and two days later Google announced an official Workspace CLI. The internet argued about the firing. Engineers noticed something else: how do you hand an agent 40 skills without drowning it?

The naive answer fails. If you paste 40 tool definitions into every request, you pay for ~12,000 tokens of schema on every single turn, and accuracy drops as the model picks the wrong tool from a wall of options. The trick that makes large skill sets work is progressive disclosure: show the agent a tiny catalog, and load a skill's full spec only when it actually needs it.