Ask HN: Where is the programming profession going?

Anecdotal experience amongst my small team:

- spending exorbitant amounts of time up front planning, surfacing every milestone, subtask, and individual change, burning tons of tokens in addition to man hours fixing minute but important mistakes or derivations despite explicit instructions, CLAUDE.md, memory subsystems, agent and skill instructions, etc.

- executing and reviewing results compared to the plan and finding even more mistakes and derivations from the plan often takes a lot of time - the relative rapidity produces a lot of output for the team which stresses our lifecycle and introduces feelings the whole process is on the verge of flying off the rails

- individual developers have different expectations, definitions of acceptable, skills/experience to detect and deal with problems, and patience. I might spend hours to days meticulously planning and executing a ticket and another guy might yolo it in 30 minutes. Other than bug escape rate and tracking review failures I’m struggling with how to track people who are “doing it wrong” let alone telling them the “right” way to do it.

- growing exhaustion, lack of ownership and confidence, frustration, and a generalized feeling of endlessly fighting your tools but in a weird way they never seem to really improve despite all your efforts to do so

- taking humans out of the loop and letting the agents be more autonomous in hopes that we’ll reduce the bottleneck and produce better results has not helped.

- I find even myself fighting (and sometimes failing) the urge to give in even though the proposed or implemented solution doesn’t feel right. Scale that across your team

- experience has not really changed despite changes in models and harnesses

- there’s a deep feeling that I’m doing something wrong and fomo since so many people in the industry boast of incredible results. I probably am, but everything I’ve read about and tried has not really moved the needle much and it’s introducing another dimension of exhaustion and frustration

overall, I don’t feel like we’re being much more productive when you factor in quality and accountability (which should be a given, but this industry increasingly overtaken by a reckless philosophy of speed over everything else). I do think it has helped parallelize tasks, produce higher quality PoCs to explore more options and do it faster, offload joyless but necessary tasks that are narrow in scope and measurable, do exploratory work and act as a generalized interactive knowledge base, and make shallow techniques, technologies, etc. that you don’t yet have experience with. Maybe I’m just missing a critical component or two in the process (formal specification, etc). Maybe it’s growing pains, or maybe there’s a looming rot. Either way job satisfaction and confidence in my work is much lower than I would have expected.

[flagged]

Perhaps solution was the wrong word for me to use here. It was intended to encompass the implementation details (abstractions, architecture, observability, etc)… All the decisions the engineers would normally make during planning and execution. Once I have that nailed down, the act of writing the code is largely mechanical.

That’s the source of my “easy” framing. It has always had the lower cognitive load in my experience. Now that I can offload the mechanical part to AI, I spend more time on the hard parts.

I still read plenty of code along the way, maybe less of it now because it’s easier to surface which parts of the code I need to read.

I think it'll probably split.

At the high end, there will always (or at least for a fairly long time) be a subset of work where AI can't do everything itself. Perhaps the framework or language is too new, or its built for hardware that's only just become possible. Perhaps the company is worried about security issues resulting from letting LLMs access its systems, or needs a human in the loop to take any liability. Perhaps the tech needs a level of performance that an AI might struggle with, and which a longtime engineer with a computer science background may be better at dealing with (like stock trading systems or something).

This will pay very well, but probably only provide enough jobs for 5-10% of the current software engineering workforce.

Then there will probably be a large number of companies where AI does most of the work, but where one or two people will probably be kept on board to guide it in the right direction/take care of issues. I suspect a lot of less technical/cutting edge companies in more 'traditional' industries will be in this camp.

Finally, a lot of work will just be outright replaced by having an LLM take care of everything. This isn't viable for FAANG companies or fortune 500 level ones, but the average mum and pop business could probably just replace its entire development team with a Claude subscription. I suspect a lot of web development agencies and lower level jobs are just going to vanish.

So, yeah. High level bespoke work where every microsecond is king? Stays mostly the same, albeit with some AI usage to handle the boring stuff.

Basic brochure/shop/CMS site development? Claude and co take over almost entirely.

Every AI coding project feels like working with legacy code. All the skills I actually use look more like managing a big messy, brittle black box codebase. It’s exactly how we would have inherited a big ugly codebase pre AI.

I think if you’re used to traditional green field development, and have never owned legacy code AI software engineering can be bewildering. If, however, you’ve been stuck maintaining legacy code, the skills serve you well In AI development.

You’re constantly looking how to tame this ball of mud. Test it end to end. Break apart bits to test. You’re chasing where the brittleness in the system exists, and you try to create a testing and feedback strategy to gain leverage on that class of problem. That in turn turns to modularization and gradually more careful organization.

I find starting out you may look at code less. As time goes on, you descend deeper and use AI to do targeted refactors to account for more constraints. You leave alone what's OK to be slop, and hone in on what needs deep care. All along adding guardrails to constrain the big ugly beast and help LLMs stay on track with their work.

The classic Michael Feathers book might be sneakily relevant: https://www.amazon.com/Working-Effectively-Legacy-Michael-Fe...