The most consequential AI development of the next few years isn't a new model—it's Skills. Here's the thesis:
- Skill will replace work—because unlike agents, Skills are controllable, auditable, and correctable.
- Millions of Skills will emerge in 2-3 years—far faster than the App Store took to reach 2 million apps.
- Unmanned companies will form—humans move from "in the loop" to "on the loop."
- Unmanned companies will outperform human-operated ones—Skills self-evolve through revenue outcomes, surpassing human-designed approaches.
Models are the power plants; Skills are the appliances. The appliance market is always 1000x larger.
Agents have a fundamental problem: unpredictability. Same prompt, different results. Fine for assistants; dealbreaker for replacing work.
Skills change this. A Skill is auditable, versioned, testable, and correctable. When it fails, you fix the Skill—not pray the model improves. Every job will be mapped into N Skills that cover its workflows.
Controllable means deployable. Deployable means it replaces your job.
Apps took 15 years to reach 2 million. Skills will hit that in 2-3 years.
A Skill is just a markdown file with instructions. Created in minutes. Better yet: AI generates Skills by watching you work once. Iterate briefly, done—automated forever.
Every job, every workflow becomes a Skill. Millions flowing, replacing inefficiency everywhere. And agents can orchestrate multiple Skills autonomously, composing entirely new workflows that never existed before.
Today's paradigm is the Human Loop: Event → Human notices → Human decides → Human acts. The agent is just a tool the human chooses to use.
With Skills, we shift to the Agent Loop: Event → Agent detects → Skill executes → Human reviews (optional or unnecessary). The human becomes an observer, not an operator.
"In the loop" means you're the bottleneck. "On the loop" means overseeing 10x throughput. Functions become unmanned when their core Skills get reliable enough. Customer support first, then sales, then operations—until the whole company runs on Skills.
Current Skills are human-written—like AlphaGo learning from human games.
But with verifiable rewards—especially revenue—Skills can self-evolve: generate variants → measure which variants make more money → keep winners → mutate → repeat.
This is AlphaZero territory. The ceiling isn't human expertise anymore. Unmanned companies optimizing Skills against revenue may achieve results that human-operated organizations simply cannot match—or even comprehend.
TLDR: Skills make AI controllable enough to replace work. Millions will emerge fast. This creates unmanned companies running on Agent Loops instead of Human Loops. Those companies, with self-evolving Skills optimized on revenue, will outperform human-operated ones.
I'm building AI agent infrastructure for enterprises in Tokyo. More thoughts on Skills and agents: @JefferyTatsuya