Software Engineering Fundamentals Matter More Than Ever

In a post published on his blog Rhonabwy on August 15, developer and author Joseph Heck argues that as AI agent harnesses make working code cheap to produce, the fundamentals of software engineering — above all how a system is put together — matter more than ever.

Crossing the “can it be done” rubicon

Heck writes that agent harnesses crossed that line in the past year, and that a harness combined with a model is a genuinely interesting power tool. The most impressive work, in his view, comes from people who are not posting about the end of the profession.

He is less convinced by the economics. From every report he has seen, the business models of the large model providers are not viable — but the capability is not going away. It is shrinking fast: open-weight models already make a beefy personal computer quite capable, not quite as effective as the frontier, but with a gap in time and capability that is not large.

“It works” is the start, not the job

Heck compares it to learning to weld in his twenties, when he quickly made things he could not lift or even get out of the shop door. The lesson transfers: how something goes together makes all the difference. With a little foresight, agentic harnesses can deliver code that works and is testable — he leans heavily on the prompt “develop with red/green TDD” — but not much above that. The seams of software — how your code works, its API, how it fits with other software — are as much art as science, resting on judgement, experience and guesses about how a system will be lived with over years.

Where the models fall short

Making software debuggable, maintainable, layered and composable remains hard, and much of that work requires extensive, thoughtful reasoning — exactly where today’s models fall short. LLMs do not reason, Heck writes; they predict, and the models are effectively compressed human knowledge. They can echo reasoning that humans encoded, but that is not the same as reasoning. He points to the paper “The Illusion of Thinking” on how badly LLMs perform at reasoning, and to research on predicting the results of actions — JEPA models, the LeWorld Model, recent talks by Yann LeCun — as a separate and fascinating direction.

Getting more out of the tools — and their limits

There are still many ways to make LLMs more effective, he says: feeding them good, concise data at the right time, and giving them deterministic validation tooling plus natural-language feedback they can use to correct themselves. What impresses him is less the prediction of text than how well models call tools and follow instructions.

That instruction-following has a dark side. He cites Simon Willison’s “lethal trifecta”: models cannot distinguish good advice from bad, and are foundationally unable to consistently prevent prompt injection. Alignment work, safety harnesses and sandboxes add barriers against the worst outcomes, but fundamental gaps remain — and something that follows instructions tirelessly without good reasoning, he writes, is nightmare fuel.

He hopes post-training and RLHF will soon reinforce reasoning about clean interfaces and maintainable design.