How Effectively Agents Write D

[jmh530]

On Friday, 6 November 2026 at 15:23:33 UTC, Peter Alexander wrote:
> [snip]
> Keen to hear any and all feedback. Also interesting in what 
> other ideas people have for how to measure D for agents.

I can't speak to measuring, but using agents is far more 
productive than programming with a chat bot.

I've been using chat bots with programming for the past year or 
so, and it's fine but it can be painful for some things. Like at 
work, there's a legacy code base that has no unit tests. So any 
time you want to evaluate it, you have to run the entire code on 
different branches and then compare outputs. With a chatbot, 
trying to add new features and making sure you don't break 
anything was a really frustrating experience.

But since roughly when GPT-6 Astra came out, I've been using 
Codex at home (for D) and GitHub Copilot at work (with that and 
GPT-6.1 Sol), and there has been a significant increase in what I 
can accomplish. The legacy code base I mentioned above, I'm 
setting it to adding a better set of unit tests. The nicest thing 
about Codex is that I can use it wherever I am. So if I'm walking 
the dog, I can check in on what it's doing. Don't need to be tied 
to a chair. That's probably the biggest plus going for it IMO.

For D, working with GPT-6 Astra I found and fixed a few bugs in 
DMD (and many more in my own code).

That being said, even though GPT-6 Astra is very smart and a 
better programmer than me, it still does some things that are 
probably either over-engineered or not the most elegant way to do 
things. And when you get a bit of a mass of code it takes a 
little bit longer to figure out what the elegant thing to do is. 
I've been working on a fairly large PR and when I try to review 
all the changes, I'm just a bit baffled at how everything fits 
together, so now trying to split it up into more manageable 
nuggets.

And I've experienced all this at work too, particularly when 
others are reviewing the code. There is a balance between moving 
quickly and doing good stuff. On my own projects, I don't have as 
much of issue with letting the agent go a bit wild and then 
regularly try to fix up or refactor. Need to take a little more 
care when working with others.