I love great software and agree that great software is evolved - not built. The whole point of building v0.1 is to figure out what's wrong and should be fixed in v0.2
At the same time, there are broadly three motions in the loop:
- thinking/discussing (what should it do)
- building (Make it do that)
- using (Seeing whether that is actually what it SHOULD do)
And then of course you repeat until you run out of time, money, patience, volition, etc. For some software there is a terminal state - it truly does exactly what it should. For most you're always reaching for it.
LLMs definitely accelerate 2 and potentially can help accelerate 1 and 3. As such cycle time can be reduced. It still may take 100 turns to get what you want, but I'd be surprised if the clock time for the 100 turns would be unchanged using AI.
> LLMs definitely accelerate 2 and potentially can help accelerate 1 and 3.
People that do think of using LLMs to accelerate 2 usually don’t think enough about 1. And from what I’ve seen, they quickly get tired of 3.
Most quality software I’ve seen usually starts from a small subset for the cycle and then incrementally add to it. LLM projects usually rush the building part and do too big of a job and it’s become cumbersome to design (sunken cost) and evaluate (too many variables). If you want to build shelters, you start with a small hut, not with a cathedral.
> Most quality software I’ve seen usually starts from a small subset for the cycle and then incrementally add to it. LLM projects usually rush the building part and do too big of a job and it’s become cumbersome to design (sunken cost) and evaluate (too many variables). If you want to build shelters, you start with a small hut, not with a cathedral.
What on earth makes you think that you can't adopt this exact approach with agentic coding??
This is like saying that you shouldn't drive a car around suburban streets, because the maximum speed it can do is in excess of 250 kph. You absolutely can (and arguably should) take agentic coding slow and steady, and build software the way you always did. Like using a car to get to the shops at 50 kph, you'll still get there much faster than if you walk.
> What on earth makes you think that you can't adopt this exact approach with agentic coding??
It can be done, but it’s not where the hype and the practice is. It’s all about number of commits, number of PR, churns in LoC, which has no bearing on software quality and usefulness.
Something like caddy[0] is just 2700+ commits over 7 years. That’s like 32 commits a month in average. Even if you apply an exponential decay (going from greenfield to mature project), the latter years would have been way peaceful. I know AI can help in some cases, but is it such a pressing need. That would been like taking a car to go somewhere less than 10m away.
Most (mature) open source projects progress a few changes at a time. Most of the time is about ensuring that nothing breaks due to the change.
I completely agree. And yes, I think the current game-ification of repositories (more stars! more PRs!) is pretty counter-productive.
I suspect that projects that run hot and fast with LLMs will be very much like driving your car at 250 kph to the shops. You might get away with it briefly, but you'll ultimately crash and burn. As with all things -- the obsession will fade soon enough, and agentic coding will become just another tool, where those who use it well will be valued. At least, so I hope. We will see, I guess.
How many cathedral does the world needs? And even if you wanted to build a new one, it’s no longer the middle age. No one is inventing UNIX all over again. Lots of hurdles have been solved. The difficulty of software is not technical or implementation related, it’s mostly about defining the problem.
At the same time, there are broadly three motions in the loop: - thinking/discussing (what should it do) - building (Make it do that) - using (Seeing whether that is actually what it SHOULD do)
And then of course you repeat until you run out of time, money, patience, volition, etc. For some software there is a terminal state - it truly does exactly what it should. For most you're always reaching for it.
LLMs definitely accelerate 2 and potentially can help accelerate 1 and 3. As such cycle time can be reduced. It still may take 100 turns to get what you want, but I'd be surprised if the clock time for the 100 turns would be unchanged using AI.