Revisiting Joel's Test - exe.dev blog
3 points by jacky
3 points by jacky
The exe.dev blog has become very very interesting to me recently as it seems a concerted campaign across multiple authors to push a very specific way of working with agents with lots of upsides promised and no downsides discussed. Whenever I've tried to work in a way the authors describe though, I've unfortunately failed.
Maybe it's just that what I'm doing is just not "good enough" or there's simply a "skill issue" but, given the product being sold by exe.dev, I think I'm justified to be a tad skeptical.
Whenever I've tried to work in a way the authors describe though, I've unfortunately failed.
As someone who does use agents, I am really amazed at all these people insisting humans should never read the code. I mean, if you do read the code, you will occasionally discover some pretty bad stuff. So I guess if you don't read the code, that's OK, because you won't ever know your agents did terrible gross things to your code, and you can somehow sleep well at night?
I've been reading other people's code for decades now. And I've also noted that some of the genuinely best CEOs know how to micromanage when they need to. Jobs once got upset at a circuit board for being ugly, even though you needed special tools to even see it. Bill Gates could talk about leap-year handling in Excel's date functions.
But so many programmers using AI seem weirdly driven to become a caricature of an ignorant, hands-off boss who has no idea how anything works. And I'm like, that has never worked in the whole history of technology, and it has indeed been one of the biggest complaints of engineers. But the moment you give some of those engineers AI minions to boss around, they attempt to become the most ignorant "boss" possible.
I don't think best practices around GenAI have solidified nearly enough that anyone can come up with a new Joel's Test for software teams.
I think the point of the original 12-point test was to transcend specific tools and remain relevant across fields and technologies. There's a reason it doesn't say, "Do you use Git?"
Vom. Another LLM acolyte saying we should just give all our money and control to the oligarchs if you want to be a "high functioning" team.
There’s some good stuff here, but there’s some stuff to quibble with. For example, while it’s great to have an agent review the code, it’s still essential to get another developer’s review in order to avoid AI-facilitated solipsism.
It seems a little early to nail these down as clear positives.
Here is my own addition to Joel's test:
Everything else can follow. Agents, should you chose to use them, can use the API and CLI tools. So can classic automation. All the rest, the merge queues, the release cycles, the testing loops... it's all fad-of-the-day.
When you can do all your work from a shell, a computer can do those steps later.
oh fuck no.
even if the joel test is dated and needs a 25yr revision, THIS AIN'T IT.
GTFO