Liminal Coding Approach
I used to write + maintain code.
Now, I half prompt/half monitor agents. But I still update my mental model of what's supposed to be happening.
In the future, I'll still be able to see what's going on1, but will rely on agents to manage codebases.
Yesterday, I wanted to grab more than a handful of screenshots: too many to do manually, and too repetitive to delegate to a stochastic agent. This belongs in code.
I one-shot prompted the coding agent to do the whole thing from soup to nuts, and watched as it chose to install puppeteer rather than playwright. Meh, ok. If a future agent decides it's worthwhile to switch, it's gonna switch. It doesn't really matter right now, and that's powerful for knowing how to use tools.
Nobody double-thinks whether their plumber uses a monkey wrench or pliers.
Actually, maybe the right analogy is, my pipes use Imperial fittings rather than metric sizes (python vs typescript). I choose whether to hire a capable journeyman or an expensive master plumber (deepseek vs codex). Plumbers make it easy to hire them. You need money and social norms, not domain knowledge.
The bits of code-craft that don't fit this future2, we should leave behind as soon as we can safely discard them.
The principle that an individual human within the team is responsible for every piece of code, will soon be outdated.
As a homeowner, I am responsible for every pipe, screw, and nail in my house. That means I have to make economic decisions on how to maintain it, to preserve the utility and value of my asset. It does not mean I need to know how the fittings work, or even the justifications for planning decisions at the time the house was built.