Ask HN: What has been your experience with Agentic Coding? | Hacker News Reader

Ask HN: What has been your experience with Agentic Coding?

7 pointsgrandimam7 comments

I have been experimenting more deeply with agentic coding, and it’s made me rethink how I approach building software.

One key difference I have noticed is the upfront cost. With agentic coding, I felt a higher upfront cost: I have to think architecture, constraints, and success criteria before the model even starts generating code. I have to externalize the mental model I normally keep in my head so the AI can operate with it.

In “precision coding,” that upfront cost is minimal but only because I carry most of the complexity mentally. All the design decisions, edge cases, and contextual assumptions live in my head as I write. Tests become more of a final validation step.

What I have realized is that agentic coding shifts my cognitive load from on-demand execution to more pre-planned execution (I am behaving more like a researcher than a hacker). My role is less about 'precisely' implementing every piece of logic and more about defining the problem space clearly enough that the agent can assemble the solution reliably.

Another observation has been that since the cost of writing code is minimal as agents are delegated to write them, there is a need for me to shift and context and also take up the QA role to evaluate the agents output.

Would love to hear your thoughts?

Look, by this time, most should already have realized that vibe coding or agentic coding is doing more harm than good. It is a great tool if you use it for debugging,code understanding , writing/refactoring a small unit of code because it speeds you up and help you learn. You're doing it right if llm helps you understand and write code faster but you must be the driver and be able to have a clear mental model of how the code works. If you just write spec and ask llm to implement and only look at the code when something goes wrong, you would quickly learn that the llm debt will grow exponentially and it is counter productive.

I have to externalize the mental model so the AI can operate with it.

Just as writing clarifies thinking, so does this. That's not a cost, that's a prudent investment.

What works: delegating non ambiguous tasks, let them happen in async, while supported by harness of preexisting automated tests, and established project conventions

What does NOT work: I have no idea how to do sth, and I hope agentic coding will solve my problem.

Think "Eisenhower matrix":

- X: Ambigous <-> Trivial

- Y: Can wait <-> Urgent

Urgent&Ambigous => Agentic Coding is useless, and an act of desperation

Can wait and at least non amibogus => Agentic Coding is perfect fit

My experience with coding by chat bot is about the same as my experiences welding by ear and driving by touch.

I'm sure there's people out there doing it but I don't feel like it's a particularly good use of my time.

It works fine for some simple tasks, but it fails at harder tasks. My struggle is to have the discipline to use it a the right time. I tend to become lazier and lazier, deploying code that I don't fully understand.

My experience is that you still have to care about some of the internal details, currently. The models are not always the best at maintaining a coherent concept for separation of concerns, a balance of automated testing, etc. Within 5k loc I had to step in and do direct refactors as the pure vibe coding approach had devolved into slop.

I'm doing a very simple frontend and backend thingy. I'm basically vibe coding my own project management tool in the way I like it. I'm currently at 4000 lines of code and am beginning to get to the threshold that I can't take technical debt anymore and would need to restructure and refactor.

So yea, fair enough.

Though, the shift that the OP describes, yea I can see that. Writing tests has become way more important. Or well, it feels more important. From a testing perspective, we should see ourselves agents too (aka bug making machines), that's why you need tests. The silly bias I always had was "but I'm writing the code! It'll be fine, I won't make bug- oh... why can't I close my modal window when I click on the x symbol?"

But yea, the apparent need for testing is definitely much more there. The need for architecting it well is also there as LLMs still seem to be a bit in tutorial land with that one. There are a few more things like that.

Reply on news.ycombinator.com