Code First, Theory Second

Originally published on my LinkedIn.

During a recent discussion on how LLMs and agents are changing programming workflows, I realised that one of the deepest shifts we're experiencing is in the theory < - > code iteration itself.

Before LLMs, we usually began with a fairly detailed mental model - a theory of how the system should work. The code was an expression of that understanding. When something broke, it meant our theory was incomplete or wrong, so we refined the understanding and rewrote the code.

In other words, understanding preceded code.

Now, with LLMs and agents, that order has been effectively reversed.

With LLMs, we often start with only a vague sense of direction. The LLM generates working code, and our job becomes to understand both the code and the problem space simultaneously. The code leads, and the theory follows.

Interestingly, this now leads to us programmers feeling as the slowest cog in the machine.

We're used to starting with understanding - where the real cognitive effort happened invisibly, before you ever wrote code. Now, since code comes first and cheaply, we're forced to build understanding afterward, and that mental effort suddenly becomes visible.

So next time you get that feeling of being the "slowest cog", remember that it's not that you've slowed down. It's just that the locus of difficulty has now moved - from construction to comprehension.

But comprehension remains as essential as ever.