After the last two weeks of using Claude Code a lot, I'm pretty sure that Anthropic's CEO predictions are correct. The future of software development (at least the majority of it) it is going to be managing agents and humans writing 10%-20% of the code.
Probably highly specific business rules code, that require a high level understanding of how subsystems interact with each other or highly performance critical path that require fine tuning with production usage data. I also think that social media / vibe coding influencer narrative is bullshit. Like any other tool you will need to master it and understand its limitations. I think of it as having 3 times the size of my engineering team with devs that code fast but might not be the most experienced ones.
Therefore, high level architecture, modularity, reducing (AI impact radius), being good a review PRs is critical. Prototyping and iterating is way cheaper now. For startups throwing away you entire / part of your codebase is cheaper now. Which can help you not being slowed down by previous design decisions while you find PMF.
That being said, the counter-argument is that if not kept in check, reaching to the point of a codebase that's unmaintainable and that nobody can understand can be achieved way faster. The "prototype becomes production" thing is way more tempting. I worry that business incentives, specially in a corporate environment accelerate the Inshitification of everything. Not due to the quality of the code being generated by LLMs but by humans cutting corners "just to satisfy management" or "the business". Highly experienced developers that have a strong voice in the table and can articulate this, slow down when necessary will make a huge impact.
Wether we like or not, managing a couple of agents, reviewing PRs, being good a translating user needs into prompts, understanding the actual user needs (not what they say they need), being good at working with cross functional teams will play a more significant role in my opinion. The stereotypical isolated, grumpy, technical dev that thinks other people code is bullshit will probably have a way more difficult time adapting.
I also worry about business pressures and the junior developer market. If in the short term hiring junior devs makes no sense, and having a few semi-senior / senior devs with a lot of agents is more cost-effective, we might reach a point where experience and knowledge is lost. It would be hard to make the generational shift and might reach the point where only AI understands the software system we use daily. I think that's a pretty bad future but will see.
There's also the impact in the overall labour market, concentration of capital, high dependency of outsourcing "thinking" and stronger dependency on AIs sold to us by a few corporations. This might be mitigated if 90% of what we need from AIs is eventually accomplished with open-source models and inference run in (relative) affordable hardware (wether local or cloud based).
Lastly, there is the identity crisis that we developers are experiencing; but that's a topic for another thread.
Probably highly specific business rules code, that require a high level understanding of how subsystems interact with each other or highly performance critical path that require fine tuning with production usage data. I also think that social media / vibe coding influencer narrative is bullshit. Like any other tool you will need to master it and understand its limitations. I think of it as having 3 times the size of my engineering team with devs that code fast but might not be the most experienced ones.
Therefore, high level architecture, modularity, reducing (AI impact radius), being good a review PRs is critical. Prototyping and iterating is way cheaper now. For startups throwing away you entire / part of your codebase is cheaper now. Which can help you not being slowed down by previous design decisions while you find PMF.
That being said, the counter-argument is that if not kept in check, reaching to the point of a codebase that's unmaintainable and that nobody can understand can be achieved way faster. The "prototype becomes production" thing is way more tempting. I worry that business incentives, specially in a corporate environment accelerate the Inshitification of everything. Not due to the quality of the code being generated by LLMs but by humans cutting corners "just to satisfy management" or "the business". Highly experienced developers that have a strong voice in the table and can articulate this, slow down when necessary will make a huge impact.
Wether we like or not, managing a couple of agents, reviewing PRs, being good a translating user needs into prompts, understanding the actual user needs (not what they say they need), being good at working with cross functional teams will play a more significant role in my opinion. The stereotypical isolated, grumpy, technical dev that thinks other people code is bullshit will probably have a way more difficult time adapting.
I also worry about business pressures and the junior developer market. If in the short term hiring junior devs makes no sense, and having a few semi-senior / senior devs with a lot of agents is more cost-effective, we might reach a point where experience and knowledge is lost. It would be hard to make the generational shift and might reach the point where only AI understands the software system we use daily. I think that's a pretty bad future but will see.
There's also the impact in the overall labour market, concentration of capital, high dependency of outsourcing "thinking" and stronger dependency on AIs sold to us by a few corporations. This might be mitigated if 90% of what we need from AIs is eventually accomplished with open-source models and inference run in (relative) affordable hardware (wether local or cloud based).
Lastly, there is the identity crisis that we developers are experiencing; but that's a topic for another thread.