It still writes like a junior dev, in that despite AI being able to get a picture of an entire repo, it's changes are typically confined to the task it's working on and will opt to duplicate logic to keep changes contained. Again, technically works, not ideal.
BUT I have had great success using AGENTS.md and becoming better at prompting to get it to not be like this.
Basic approach in AGENTS.md: don't code defensively, yada yada, we have a validation layer at X, no need to check for anything behind that layer. Works well.
An approach I've found helpful when prompting: What would be the best architecture for this change? If you say "do X" it'll tend to just do the hackiest, shortest path thing. If you say, "what's the best way to do X?" it will think more holistically.
That said, who knows, maybe when it's PHP it just really wants to hack ;-)
(Also, yes, you still need to review the code -- it will still do stupid things, so you can't just be pure hands off w/o ending up with quality degredations. The same is true of humans too though in my experience...)
Since you’re not in a unique position, I can confidently state that your comparison of LLMs to jr developers seems unfounded. Today, LLMs produce code that is superior to junior developer code by an order of magnitude.
Notably, they demonstrate consistent syntax, clear separation of concerns, strong test coverage, organizational rigor, idiomatic API usage, and the ability to generate and maintain documentation, among other measurable qualities.
LLMs generally operate at a staff engineer level for a number of languages and ecosystems (including polyglot projects).
Comparing an LLM to a senior developer is an absolute joke.
2. Are you referring to without having a compiler or LSP check it? Although even then, the recent LLMs I've used still frequently get syntax right, whereas I'd expect juniors are often using a LSP or compiler to catch mistakes while writing code?
who cares about syntax? who cares about iteration? what I care about are _results_, which they can produce at the end. do you check your human colleagues how many iterations they do before committing/showing their work to anybody? no. why should you set such a bar then for your LLM?
We have many folks (not engineers) at our company using LLMs to open PRs, and every one of these PRs has profound architectural design problems.
This is a critique of scale, moving the goalpost.
There is serious magic happening in the construction of model context.
> The python visualizer tool has been basically written by vibe-coding. I know more about analog filters -- and that's not saying much -- than I do about python. It started out as my typical "google and do the monkey-see-monkey-do" kind of programming, but then I cut out the middle-man -- me -- and just used Google Antigravity to do the audio sample visualizer.