Eating cake

Why you need to put down the teaspoon

I’ll be honest. I’m partial to a good cake. And, for many years, creating software has been a bit like eating cake. But while the cake is tasty, it’s been hard to eat because we’ve only been able to use a tiny teaspoon.

AI changes that. We’ve now got a massive spoon. We can guzzle cake like never before.

Yet, so often when I talk to developers, it’s clear they are still nibbling at the cake with their new massive spoons. Or desperately clinging to their teaspoons. Earlier today someone was explaining how they carefully approved each request from Github Copilot. A few weeks ago a team told me they didn’t want the AI to generate too much code because they would have to review each line by hand. Following the established process was something they held dear.

Your new spoon

I get it. These are the processes that have kept us safe for decades. It’s how we do software engineering.

But now?

Now I have fifteen desktops with over ten projects on the go. Each has at least two instances of Claude or Codex running. They’re not all active at once - that’s currently beyond me - but they all make progress every day. Some are running long-running code-coverage improvements and have been going for days; others are more interactive.

I don’t write code anymore. Instead I have ideas. I describe what I want. I sketch architectures. I smell when something’s wrong and ask questions until I understand why. I context-switch constantly, keeping the plates spinning.

It’s product management and architecture and team leadership, all at once. The skills I’ve gained over the decades - a deep knowledge of C, debugging instincts, understanding the detailed software development process - they’re not useless, but they’re no longer the point.

The harsh reality is that many of the skills we associated with software engineering are rapidly becoming obsolete.

It’s the process, stupid

It’s not just skills; it’s the processes as well. When I started in the industry over 30 years ago many software engineering processes were still in their infancy. We were all still learning.

But over the intervening decades processes have started to calcify. Take code review. It’s held up as a gold standard - a thing we must do otherwise unspeakably terrible things will happen. Yet it wasn’t always this way - I know of plenty of production code that was shipped without code-review (probably best not ask how I know…). And in those days, when your code shipped on floppy disk and patch cycles were measured in years, the consequences of bugs were arguably worse than today.

Ultimately code-review is primarily about quality. And if we can achieve quality in a different way, then perhaps we can replace human code-review with a combination of AI review and improved testing.

Or consider testing. In the past a human tester would find a bug and raise a ticket. It’d get triaged and eventually a developer might fix it. In this new world, a test agent finds a bug and then just goes and fixes it. All the diags are right there. No need for tickets. No need for triage - it’s easier to just fix everything.

Or product management. Development used to be expensive, so we spent a lot of time working out which features to implement. Now? Maybe we should just implement them all. At a minimum we need to be careful how much human time is spent to save a negligible amount of AI tokens.

This is a completely different way of doing software engineering. Our processes need to change. Trying to fit AI into the old way of doing things means we’re only able to use a fraction of the capacity of our new spoon.

And so?

So is this all upside? Obviously not. The models aren't perfect yet. They require oversight. Some mistakes are subtle. Some less so. Here's a bloopers reel from the past few weeks…

3D (or even 2.5D) is a struggle - this is meant to be a racing game…

Running multiple agents in parallel risks fights, with the models reverting each other’s changes - and then blaming the linter…

Gemini refuses to believe we’re in 2026…

But often makes me smile…

Things go wrong - this is Formula One GP after Claude optimized the video support…

And how is your Spanish?

But then things like this happen. And it turns out that letting Claude use a massive spoon has created a working 386 emulator with embedded DOS and sound.

And I realise we’re living in the future.