Agents Can Build. What Is Left to Learn?
Agents may replace some lower-level knowledge work, especially in software. They can already write code, search documentation, and assemble an application. But they cannot replace the human desire to learn or be entertained, whether that desire comes from progress, curiosity, or simply passing the time.
If agents eventually handle much of the work, what will people do? Beyond material needs, we still have mental needs. We want to understand, express, appreciate, and create.
I think of knowledge and art as having their own input and output:
| Area | Input | Output |
|---|---|---|
| Knowledge | learning | teaching |
| Art | appreciation | creation |
An agent can assist with every box in that table. It cannot make me want to enter one.
Learning in the age of generated software
Software and electronics are subjects I want to keep learning even after I retire. That makes them worth taking seriously now. The rise of vibe-coded software does not reduce my interest in fundamentals. It increases it.
People worry that using an agent means they will stop learning. The answer is not to reject the agent. It is to choose more carefully what deserves to be learned. I want to understand the foundations underneath agents and software systems: how an agent loop works, how a browser works, how a terminal works, how a model is trained, and how an isolated runtime works.
There is no shortage of generated explanation. In fact, that may become part of the problem. More documentation and websites are being written by agents for other agents. A human who wants to understand the system may need to work harder to find material that was tested, digested, and organized by someone who genuinely learned it.
The agent can still be a powerful assistant. It can answer questions, trace code, prepare examples, and shorten the feedback loop. But the learner needs a path through the material and a reason to follow it.
What good learning material looks like
Projects such as Build Your Own X and CodeCrafters make systems understandable by asking people to build smaller versions themselves. Eureka Labs and Andrej Karpathy’s Zero to Hero explore how expert-created material and AI assistance can work together.1
The subjects I keep returning to are similarly foundational:
- Ghostty as a path into terminal and native application infrastructure
- browsers as a path into rendering, networking, sandboxing, and the web platform
- virtual machines and microVMs as a path into agent runtime isolation
- llm.c and llama.cpp as paths into model training and inference
- Pi as a path into a small, understandable agent loop
- Omarchy as an example of rethinking the computer around a newer generation of tools
These are not easy subjects. That is part of their value. They sit below the fast-moving product layer, so understanding them creates knowledge that can transfer to whatever comes next.
Turning knowledge into ability
Collecting information is not the same as learning. Knowledge becomes useful only after I process it, organize it, and apply it to a real problem. Original work comes less from remembering one specific solution and more from developing an abstract ability that can solve many specific problems.
This is also why I want to rebuild my habit of taking notes. I sometimes encounter the same problem twice and realize that I have forgotten the earlier solution. A useful learning system should not become a drawer full of facts. It should help turn repeated experience into an ability I can use again.
In the AI era, the people closest to real problems may have the strongest advantage. An agent has broad knowledge, but the person working on the actual system sees the constraints, failures, and consequences. Learning should bring us closer to that reality rather than farther away from it.
Why I want to build here
I like learning, digging down to first principles, pulling a complicated subject apart, and putting it back together in a form that can be taught. I also like teaching and the respect that comes from helping someone understand. As a dad, I want to give my children better ways to learn. As a software engineer, I have enough foundation to begin with software and enough end-to-end building experience to turn material into a working product.
This is a difficult area because good teaching takes interest, patience, and persistence. Many people are busy following the latest thing and shipping another thin application around it. Fewer people want to spend months understanding the layer below.
That gap is exactly what interests me. If agents make software easier to produce, then carefully learned and carefully taught fundamentals may become more valuable, not less.