Introduction

AI developments (and especially in the agentic AI assisted development space) are still moving very fast even 3 years after the release of ChatGPT.

I heavily use AI both at my day job and in hobby projects. I'm on X almost daily, can recommend LinkedIn as a good source of information, and am enamored by Simon Willisons blog.

But even I still feel like I'm missing things due to the sheer volume and velocity.

Not everything is worth your attention though, and sometimes you have to make choices for your sanity (for instance, I haven't touched OpenClaw yet with a ten-foot-long pole).

I did however find two good videos I'd like to share with you that paint a broader picture of the more interesting recent developments.

Both more or less arrive at the same conclusion:

IMPORTANT

We've crossed a threshold where AI agents can handle long-running coding tasks mostly unsupervised. Handwriting code and overly babysitting the agent is mostly holding you back, so you need to upskill and drop the emotional bond you have with your code to keep up.

They also give some helpful tips on how to position and sell yourself as a Software Engineer now that code is extremely cheap. Long story short: systems thinking and domain expertise has become even more valuable than before.

Video 1: Jo Van Eyck — You are not ready

Jo shows us some nice new tooling in the agentic AI space, and talks about the 'AI maturity ladder'. (Not to toot my own horn here, but I feel like I'm pretty high up on the ladder😅).

A nice point he makes: due to the sheer volume of code Claude Code can spit out, it's hard to still have (emotional) ownership of the code.

Back when you handwrote code, you had an emotional bond to it and obviously a deep understanding (until you had to revisit it a year later 😅).

Jo proposes to gain back some of this ownership by doing a refactoring pass after the agent is done developing. Even though the agents produce decent enough code now (not quite up to my handwritten standards, but approaching it), a refactoring pass gives you some of that ownership back.

Funnily enough, I used to do exactly that when I was still handwriting code and needed to get accustomed to a new codebase. Learning by refactoring. Nice how it still holds up even when the paradigm has shifted.

Highly recommend.

Video 2: Nate B Jones — The Five Levels of AI Coding

This video is a bit more on the sensationalist side, but I really liked the five levels framework — which is kind of similar to Jo's AI maturity ladder:

Level 0 — Spicy autocomplete (AI suggests next line, you accept/reject)
Level 1 — Coding intern (hand AI a discrete task, review everything)
Level 2 — Junior developer (multi-file changes, you still read all code)
Level 3 — Developer as manager (you direct AI, review at PR level)
Level 4 — Developer as PM (write spec, leave, check if tests pass)
Level 5 — Dark factory (spec in, software out, no human writes or reviews code)

So we've moved from basic code completion (level 0) towards level 5 where the AI basically does all the work. Some software shops seem to exist succesfully at level 5.

IMPORTANT

The creed of these 'level 5' software shops is that no code should be written by humans, let alone reviewed by humans.

The main example he gives is StrongDM.

I remain very skeptical, but it's hard to argue with examples that make money. After all, our job was always to deliver working products, by the fastest means possible.

For now I'm happy and productive at Level 3.