The compounding effect is going to be big in agent development, building agents, and orchestrating large models, and it’ll get more visible over the next three to five years.
I use AI tools heavily: Claude, Codex, Antigravity, and others. I’ve burned through more than a billion tokens. Everything here comes from hands-on experience: using agents to write code, build tools, draft docs, generate images, and produce videos.
Why AI doesn’t know how to use AI
AI learned from code repositories, human writing, and art, and it built up real knowledge and capability from that. But across all of human history, there’s zero data on how AI uses AI. It’s starting from nothing. That produces a strange result: AI is good at writing boilerplate and production code, but when it comes to writing agent orchestration logic or agent prompts, it’s bad at it. Whether it’s Claude Opus 4.8 or GPT-5, the prompts they generate are mostly off. They look fine at first glance, but once you actually run them, you find all kinds of problems. This is a hard problem for AI to fix on its own, because it has never learned how to use AI. It can only fake it with Skills, spec docs, and similar aids, and that’s nowhere near as good as the codebases baked into its original training data.
Training takes time
AI training doesn’t just take time, it takes massive resources, electricity, and high-quality data. And the source of that data is the prompts all of us write when we use AI. We’re still at an early stage. Today’s Skills, the earlier MCP wave, the “magic spell” prompt tricks before that — they’re all just prompt engineering evolving. In other words, we’re still in the early stage of learning how to steer AI, let alone AI steering itself. This window might last three years, or maybe just one. Either way, it exists, and the demand it creates is real.
Why the compounding effect is so big
First, every upgrade cycle and every new architecture in AI comes fast. But because each leap is fast, there’s never a deep barrier that builds up. Compare it to Linux: if you want to upgrade it or contribute to the kernel, the amount of knowledge you need is huge. But look closely at each iteration of AI usage techniques and there’s not much to it, mostly scripts and prompt tweaks. Once you learn the basics of using AI and pick the right tools, you can get your head around what MCP, Skills, OpenClaw and the rest actually are in about the time it takes to eat a meal. If you understand today what AI is and what an agent is, every future iteration is just incremental knowledge, and with AI’s help you pick it up fast.
My take: the compounding effect of agents comes down to the skill of using AI. Some people use image-generation AI to make good visuals. Others use AI to build projects like Hermes or OpenClaw. Both point to the same thing: AI amplifies what a person can already do.