The AI Coding Paradox: Fewer Engineers, More Engineering

In 1986, Fred Brooks published No Silver Bullet, arguing that while technology could simplify coding, it couldn't eliminate the genuinely difficult parts of software engineering: understanding requirements, designing systems and managing complexity.

For 40 years, that argument held up pretty well. I'm beginning to think AI is challenging it.

I'd argue we've reached the point where a single experienced engineer, orchestrating AI, can accomplish work that previously required a traditional development team.

And it's not just coding.

Requirements, architecture, design, risk analysis, product security, documentation and verification can increasingly be automated or greatly accelerated. These aren't simply the tedious parts of software development. AI is taking on much of what Brooks considered essential complexity.

Software engineering is an extremely cognitively demanding discipline. It requires abstract reasoning, problem-solving, creativity and the ability to manage extraordinarily complex systems.

If AI can increasingly perform this work at or beyond the level of experienced engineers, why would we assume chemistry, biology, finance, program management or other knowledge-intensive professions are somehow beyond its reach?

I think we're fooling ourselves if we believe this transformation will be limited to software development.

But there's an important distinction between generating software that appears to work and establishing that it's correct.

Neither humans nor AI can exhaustively test every possible behavior of sufficiently complex software. Verification has always been about managing risk, not proving perfection.

As AI takes on more of the development lifecycle, I believe independent verification becomes even more important. Not simply testing whether software meets its requirements, but questioning whether those requirements are complete in the first place.

If AI develops and verifies software using the same flawed assumptions, we could end up with a system that passes every test but is fundamentally wrong.

And here's the paradox.

Organizations will naturally see these productivity gains and question how many engineers they actually need. But the people who understand the architecture, failure modes and limitations of verification are precisely the people best positioned to recognize when AI gets something wrong.

I don't believe AI will always need that human expertise. In fact, I suspect it will eventually outperform us at much of this work.

But there's a difference between recognizing where the technology is heading and assuming it's already there.

AI may eventually make much of today's engineering expertise unnecessary. My concern is that organizations will decide it already has.