Post

My exposure to agentic coding has been limited but in the time I used it at a freelance client I did not see the kind of leaps in productivity others describe. The code was sometimes useful, sometimes useless, and always required enough supervision to offset the time savings.

That said, for small bits of assistance or getting a task started, I found AI was a useful little tool. Developers do a lot of copy-pasting code from online, and a language model is well suited to doing a more flexible version of that, mapping human language inputs to code outputs.

As a non-coder, whenever I see coders talking about "oh it's really helpful to help me search a library and pull exactly what function I need" I'm baffled by the fact that these code libraries were apparently not well organized/not easily searchable to begin with??? Baffles my librarian mind

They're often quite well organized actually. Before LLMs, editor software had already built robust auto completion functionality using computer-readable documentation. What's harder to find is a code sample showing how to combine various functionality to match your precise use case.

Ah so these folks might just be dumb when it comes to searching, which actually doesn't surprise me (again speaking from my librarian background lol) If there's no example to work off of, then how is AI a help during that part? Does it just combine things more quickly for you to experiment with?

At its best it's quickly synthesizing things I would have been able to find myself, but faster and with less frustration. A nice thing about coding is you can write code to verify that the code the AI gave you actually works (a good practice even pre-LLMs), so I don't have to trust the output

But this is why I'm so appalled by all the vibe-coding and agentic coding stuff. It gets stuff wrong a decent amount of the time and I wouldn't want an imprecise tool like that running my entire system

The useful cut is which steps get a check the model cannot talk past. Synthesis for a one-off snippet is fine. An unattended loop over anything that looks like prod is where the imprecision stacks, because each wrong branch still narrates confidently.