Four Years Into AI, I Feel Like a Builder Again
Nearly four years after ChatGPT launched, I'm more hands-on than ever. Here is what changed my mind about guardrails, code reviews and writing briefs.
Andrew Yardley, engineering leader, hands-on with AI.CareerMore posts
On 30 November it will be four years since ChatGPT launched. I installed it on day one and it has barely left my machine since. There are very few days when I haven't used some form of AI, and I had been talking about AI and image recognition for a few years before it arrived.
I've written before about the tools. This time I want to write about where I've landed, because my thinking has moved a long way in the last couple of months. I may well be wrong about some of it. I'm sharing it anyway.
Nobody is opting out any more
Four years on, even the organisations that started out hardline are using AI in some shape or form. It is forcing the adoption. The deepest sceptics are finding that standing still means being left behind, and most of them can see it.
In the teams I lead I can see adoption develop in levels, and I can see people's progression.
"It only takes a couple of weeks"
An engineer told me this week that people pick AI up in a couple of weeks. I strongly disagree. Nobody gets efficient that quickly, because the efficiency comes from exploring, getting poor results, and adapting.
We are also a million miles from prompting now. The work today is skills, agents and the structure around them. That is daunting if you are starting from zero. It may be that AI will eventually handle all of that for us, and we're only doing it this way because everyone is still finding their feet. Until then, the learning curve is real and it deserves respect.
I see developers who feel they have maximised AI because they are good at prompting. I'm encouraging them towards skills, knowing that skills are only the next stepping stone.
Builder, leader, builder again
For the first half of my career I was very hands-on. Then I moved through team lead, head of engineering and engineering management, and leant into the people side: motivating, coordinating, organising delivery. I enjoy it. But the hands-on time dwindles. With a young family, there was never much left over.
AI has reversed that. I'm more hands-on than ever, because it lets me fast-track things I would never have found the time for. For a while I went through the motions of worrying about what it meant for my own career. I think I've turned a corner on that. Now I'm leaning in, having fun, and seeing what we can achieve. The more I do, the more I automate, and the more I feel like a builder again.
Much of that has come from chipping away at skills bit by bit, often without noticing, and learning from the varying quality of what came out.
Your agents are your guardrails
For a long time I only associated guardrails with automated tests, a security scan, or a review that had to pass before anything reached production. Permissions and access limits matter too. But the penny dropped over the last month or two: your agents and skills are guardrails as well.
Build specialists. One for CSS, one for security, one for UX, one for checking content, one for accessibility, including keyboard control and colour contrast for people with visual impairments. Put them at the end of the chain as the broker that decides whether work moves forward. It changes the mindset from "how do I write this code" to "what mini team does this piece of work need to maximise the quality?"
What I think about code review
This is where I'm most likely to be challenged, so I'll be plain about it.
I watched a talk at a leadership conference arguing that even with pull request reviews, a high proportion of code still shipped with bugs. I don't believe every figure I heard, and I haven't verified them, so I won't quote any. But it matches something I've suspected for a while: review catches less than we like to think. The teaching value is fading too, because people are increasingly using AI to write the code. What's left can feel like a bit of theatre, with a lot of nitpicking.
I can also see the practical problem. The volume of AI-generated code is so high that the people have become the bottleneck. The code was written by AI and checked by AI, so why are we getting in the way?
Nobody I know has measured the error rate with proper guardrails, specialist agents and orchestration in place. My bet is that it drops, and that well-built guardrails do a more comprehensive job than a person reading a diff. That said, humans in the loop still matter. AI makes fewer mistakes than it used to, but it still makes them. Of course, humans get the wrong end of the stick in reviews too.
Code we may not be able to read
I've believed this for a long time and have probably said it to colleagues years ago. Languages will become more efficient, and they will be made efficient with AI. The result may not be easily maintainable by humans, and reading it back may feel odd.
Machines write differently to how we do. So what problem are we trying to fix by making AI write to our conventions? I know there are strong arguments against, and I know I'm on one side of this and drifting further towards it.
A brief beats a prompt
The headline for me is that the way we have always done development needs flipping on its head. The processes we use now don't suit this way of working.
It is more product-based than it used to be. It's about how you think and what you really want to achieve. A good, solid brief gives the AI all the context it needs, and better context gives better results. Then you need the right guardrails and agents set up to follow that brief.
Proving it, not pitching it
Some people will resist, and a leader's instinct is to find the right words to persuade them. I don't think words are enough. I have to prove it. That is a new leadership challenge, and I don't mind it, because I'm hands-on with AI every day. I'm building proofs of concept and using the evidence rather than the argument.
To be an engineer or a leader in this space, you have to keep on top of how the development cycle is changing. I'm having fun with it. If I turn out to be wrong about some of this, I'll write about that too.