Beyond Vibe Coding
Beyond Vibe Coding
Is Coding Not Solved ... Yet?
0:00
-1:02:52

Is Coding Not Solved ... Yet?

Thoughts on #S2.20 Coding Is Not Solved: Accountability, Reliability and the Agentic Gap - Alex Ewerlöf

When I first read Alex’s article Coding is NOT solved, I was skeptical. Sebastian and I talk to people every week who ship more and more code with agents. Some of them barely read it anymore. So the title felt like a contrarian take.

At the same time, I knew Alex. He has been writing code for more than 27 years and spent most of his career in reliability engineering. If he says something like this, there is a solid argument behind it. So I was curious about the conversation. It did not disappoint.

It depends on the risk profile

The most important thing I learned: how far you can go with AI in building and running software depends on the risk profile of the application.

Alex gave a simple example. He built a small timer as a Chrome extension. He never wrote or read a single line of it. It runs in a sandbox, it has no access to private data, and if it breaks, nobody gets hurt. For that kind of software, relying fully on AI is completely fine.

Now compare this to software that handles payments, health data or anything mission critical. There, hoping that a harness with many feedback loops will get the quality where it needs to be is probably not enough. LLMs find an answer, not necessarily the best one. And Alex reminded us that in reliability engineering we talk about 99.9…% and more. Benchmarks are still far away from that.

I think he has a point. At the same time, many applications are simple CRUD apps. A database, a few GET and POST requests, that’s it. For those - independent from the risk profile - AI can probably carry most of the work there. Maybe this even changes how we design system landscapes: keep the critical core small and well understood, and let AI take over more in the low-risk parts around it.

Alex used a nice picture for this. In the future, software might look like cars: Volkswagen for most use cases, cheap and solid, built with a lot of AI. And Rolls-Royce for the cases where the risk does not allow AI to go full throttle, with much more human involvement.

For sure, humans are not deterministic either. Ask ten developers to solve a problem and you get ten solutions. If the system meets the functional and non-functional requirements, why does it matter who wrote the code? Alex’s answer was: It depends on the risk. And that brings me to the second point.

AI can’t be accountable

What really resonated with me is that engineering will not go away. Understanding non-functional requirements like reliability, security and performance becomes even more important.

Because in the end, AI can’t be accountable. Only people can.

Alex describes ownership as three things: Knowledge, mandate and accountability. You understand the problem, you have the power to decide, and when things go wrong, you are the one who gets called. An AI does not care about being on call. It does not care about a performance review or a GDPR fine. A human does. Someone in the company is accountable for what goes into production, whether a person or an agent wrote it.

… yet

For a certain class of applications, coding is not solved. But I would add a “yet”.

Agentic engineering is a very young field and the speed at which it develops is impressive. Alex himself said that his view is based on today’s capabilities. Two things would change his mind: models that can reason in abstract terms and models that can learn in real time. If that happens, we need to have this conversation again.

So we will. Not too far from now, I hope.

Thanks for joining us Alex Ewerlöf!

—

Links:

Discussion about this episode

User's avatar

Ready for more?