What FSD Teaches You About LLM Coding Agents

I've been driving with Tesla's Full Self-Driving for a while now, and I've been building software with LLM coding agents for about as long. The two experiences rhyme more than I expected. The biggest lesson from both is simple: use it enough and you'll feel where the fringes are.

On the highway, FSD is remarkable. It merges, changes lanes, handles stop-and-go traffic, and most of the time you forget how much work it's doing. Then you pull into a busy parking lot. It hesitates, picks a strange line, or creeps along in a way no person would. Parking is where it still can't handle things as well as you'd hope.

Coding agents in early 2026 felt the same way. Ask for a web app and within minutes you had something that looked finished: a clean layout, nice typography, a dashboard that would impress in a demo. Then you clicked around. The Save button didn't save. A filter changed the URL but not the results. A form posted to an endpoint that didn't exist. The surface was polished; some of the wiring underneath was missing.

The lesson in both cases isn't "don't trust it," and it isn't "trust it completely." It's to learn its shape. After a few weeks with FSD, you know when to keep your hands near the wheel: parking lots, construction zones, odd merges. After a few weeks with an agent, you know where to look: buttons that call the backend, edge cases in forms, anything that depends on state carried across pages. You stop supervising everything evenly and start supervising the fringes.

That's why manual testing still mattered so much in early 2026. Tests an agent wrote for itself could pass while the app was broken, because the agent tested what it built, not what you meant. Ten minutes of clicking through the real thing, the way a user would, caught problems a green test suite missed. You had to iterate: click, find the dead button, describe it, let the agent fix it, click again.

By late 2026 that's become much less common. Agents now run the app themselves, click through it in a browser, look at screenshots, and fix what they find before handing the work back. I still spot-check, but I find dead buttons far less often than I did a year ago. FSD keeps improving the same way, release by release. The fringes move outward.

And because the fringes move, your mental map has to move with them. A few habits have served me well with both. Use it every day; it's the only way to build the map. Verify at the edges, not everywhere. And re-test your assumptions after every upgrade, because what failed last month may work now, and a new release can bring new rough edges you haven't met yet.

This is most of what I help teams with. The question is rarely whether the model is capable. It's where its fringes are for your particular workflow, and how to put a person, a test, or a check right at those spots, so you get the highway miles for free and stay alert in the parking lot.