As a preface, I use Claude in my day-to-day work as I've said before. I enjoy using it and exploring other models, local LLMs, etc. but there's something strange going on in the software development space that I refuse to believe is real.
What is happening?
In the past few weeks I've seen a huge increase in posts across Reddit, Twitter, and other hellholes ranging from people talking about managing hordes of agents to "We run the whole company without having looked at the code, agents write it, agents review it, what could go wrong?". Maybe I come from a place where it actually matters to get logic correct because otherwise we do harm to our customers that we are legally required to fix at our cost, but these are absolutely insane takes. I can understand the pushes to develop faster with these tools, but as I said in an earlier post, the bottleneck was never the coding, but knowing and understanding what to build, so is everyone just throwing together any possible idea whether it makes sense for the product or not? In a greenfield project, sure, it's easy to come up with things that need to be done quickly but as a project matures that becomes harder to do without tossing out the existing architecture and behavior. Then here's another crazy one, "We don't need unit tests, they only enforce slop behavior", which....what? Are people really relying only on the agent's memory to enforce behavior? I get that there's probably plenty of mostly useless unit tests that aren't really testing anything and sure get rid of those, but if I'm testing actual logic, why would anyone want that?
My take
Tech bros be crazy, non-tech tech larping bros be even crazier. I'm glad that AI has enabled so many people to be able to build the software they want, especially for themselves. As an employed software engineer though, we (you, us, them, whatever) are still responsible for the code that's put out into the market. My career has been in healthcare and finance so maybe that's why I care so much. Maybe it's because I take some pride in my work and check it over before merging. I'm sure the tech bros will say I'm doing it wrong, but that's what I'm paid to do, not to blindly trust that the person or thing I've asked to do it has done it perfectly. Also, how do you have so much work in detailed, ready-to-go tickets that you can do this? How do you have the bandwidth to test them? Oh... right, you don't. Or you don't care about testing and "users are the testers" (thanks Microsoft).
Why I'm wrong (in some people's eyes)
I'll fully admit that my expectations probably don't align with a lot of the industry's. In a lot of places, getting 80% of it right is usually enough and the edge cases fall through to either support or workarounds. While I probably never get to 100%, I like to know what gaps there are and how big of a gap it is. I also work primarily in financial contexts so getting it 80% right can sometimes mean not getting a person's deposit / withdrawal correct and screw up money movement, people get mad over that. Personally, I believe there's a lot of personal responsibility that should go into each person's work and if you believe that you're responsible for your team's work (as you should), then the same goes for the team of agents you orchestrate. We see this responsibility in other engineering professions, but for whatever reason haven't applied it to software engineering and I believe it's one of the main reasons why software engineering isn't a certified professional engineering thing in the US.
What matters
At the end of the day, I know that the presentations on social media don't reflect the real world 99 times out of 100. What really matters is what you build. You built something to watch those 50 agents do trival tasks? Cool, show it off. You won't convince me that you know what those 50 are doing, but if you're happy, I'm happy for you. If you're actually doing this for production code and have no idea what's going into what your users are doing? Man, I don't know. For something that you built because you wanted it and others happen to use it? Sure, that's fine I guess, use at your own peril. For something you're selling and your users expect it to behave a certain way, not cause harm, etc.? You have some responsibility there and have to know at least the basics of what it's doing. Don't know how to read code? There's never been a better time to learn, hell, the same AI will teach you and walk you through your own app while it does it.
I'm sure I'd get comments like this if I was more popular, "Well I'm just doing what my boss told me to do, we have to be more productive!", and to that... well sure, follow orders, there definitely haven't been people in the wrong for doing that in the past...or currently...politics aside, that's a horrible mindset to be in. Sure the market sucks right now, but bending over when faced with that adversity isn't going to make it better. I guess it's a hard time to have integrity like that, but I think it makes me better so I'll continue to push it.
Maybe I'm the one that's wrong, maybe I'm doing it wrong, but I'm doing it the way that works for me and maybe that's why I think these posts are crazy.