[thedevilslawyer]

LLMs are plenty innovative and generate good quality code by default, and great quality code when directed well.

If you're not seeing this, at best you're probably unable to direct them or use them well.

FWIW, if you don't believe the above, I challenge you to put up a quick git repo, where you are unable to get the deserved quality out, and we can quickly show you how the same quality is available via SOTA agents, within a fraction of hand-coded time.

[larsfaye]

I do agree with this. I was able to get exactly what I needed even out of GPT3.5. If you put enough parameters and examples, along with a real solid system prompt and (if you can) proper temperature and topK/topP, there's no reason they can't basically function like "smart typing assistants". The issue is that its a sliding scale of ambiguity to chaos. The more ambiguity, the more the LLM fills that in. And it can be very difficult and time consuming to know when you're being ambiguous or not (you don't know what you don't know, OR, you can't track what you aren't tracking).

Depending on the task, it can sometimes be just as arduous to produce enough guidance and guardrails to get the LLM to output exactly what you need that you can trust without issue or extensive review than it is to write it yourself and use the LLM just for ad-hoc generation. It's a constant balance and an endless amount of micro-decisions, honestly, but it's pretty essential to stay engaged and not YOLO with agents the way so many are. Most of my interactions with models these days are done in pseudo-code.

[abalashov]

Great article! I use LLMs 99.9% in the chat box, to bounce ideas around, ask research and reference questions, and so forth. However, I have sworn off 'agents' entirely for serious code, for reasons your article captures almost totally and perfectly.

I'll still use 'agents' for throwaway tasks--mostly with local models--including tasks where some sort of ad-hoc code generation is in the critical path (e.g. scraping data).