Edward Wang

Against Agentic Coding, For Collaborative Coding

The problem

Agentic coding is bad for 3 reasons:

  1. you lose focus
  2. you lose understanding of the code
  3. the AI subtly introduces bugs and conceptual errors

We’ll start with 1:

Letting the AI write the code for you (as opposed to typing it yourself) causes you to context switch out. So you say “implement this feature for me…” bla bla bla, you go into plan mode and whatever.

Then you switch to youtube and start watching shorts.

Constantly thrashing your context like this makes you really tired, and it gets unengaging fast.

Point 2:

Finally, point 3: All of this combines to make it a lot more difficult to verify the code that the AI generates.

You might go into it with a P=NP idea where you say “it’s easier to verify code than to come up with it”

So what’s the solution?

The solution is to use the chat.com or claude.ai website, and manually type the code that the AI produces.

Furthermore, this is more of a collaborative relationship. You and the ai work together. It provides the domain knowledge, you provide the critical thinking.

Like fundamentally the chatbot interface is collaborative. The interaction is literally modelled like a text conversation in discord or slack or whatever.

And it’s a much tighter feedback loop between you and the AI. Whereas in agentic coding, there isn’t a clear time where you can step in and help the ai.

Similarly the ai doesn’t really help you understand. You have to wait till it’s done the whole thing to read it. Which makes it harder to verify correctness, and you also wasted a lot of compute if it spent time working on a wrong spec.

The strong points of AI are it’s crazy wide domain knowledge. The weakness is reasoning.

The strong points of human intelligence is reasoning.

By working together, you can produce better code and a more complete model of the problem.

What about “vibe coding”?

There’s this idea in silicon valley that it’s okay to not understand the code. And that it’s okay to write bugs, as long as the AI is fixes the bugs fast enough.

I think the anthropic compiler experiment showed that this isn’t true. Eventually the agents will get to a point where they can’t make any progress because everytime they “fix” something, they break something else. The problem here is that they fundamentally modelled the problem wrong.

In my opinion, programming is not about writing code, but developing understanding about problems and creating the correct model.