the future of software development

software development is one of the most impacted fields by ai. i want to think about how it might change in the next 5 to 10 years.

to answer that question, i think it is helpful to just take a look back at traditional programming, and how it has changed just this past year.

traditionally, startups would raise a substantial amount of capital to hire a team of engineers, and spend months (or years) developing an idea. the process would rely on highly skilled engineers who have strong experience, and those with the ability to adapt and learn quickly using resources available to them. implementing a web framework like FastAPI would involve actually reading their documentation, or perhaps looking at a YouTube tutorial, or even spelunking through open-source projects on GitHub.

things have totally changed with LLMs (obviously). i remember around the time when i was recruiting for my new grad software engineering role, ChatGPT 3.0 was good, but it still was not the best tool to one shot a LeetCode hard. it was excellent for guiding you along the right path, but not the most reliable tool for crushing questions.

it was really good for bouncing off ideas, but never for fully executing a coding task flawlessly in one shot; even if it did it quite well, it wasn't normally the most optimal way of doing it, and required multiple tweaks.

that is what my background and past experience has been before i started working full time in august 2024. at this point in time, the models have progressed a lot more to the point i was able to practically learn anything new, with good code snippets to help me understand.

i would say when Sonnet 3 came out in early 2025, is when things became tremendously easy to build. with no prior production-grade infrastructure experience, i was deploying things to AWS and using their full suite of tools in a matter of weeks, instead of the traditional way of spending a couple years using it at a company until i can add 'AWS' as a 'Technical Skill' on my resume.

the executional gap has shrunk so much that if you are a semi-competent engineer who can identify right from wrong from the LLMs, you can build almost anything. the number of times you will need to distinguish between right and wrong from the LLMs is also greatly shrinking, and shouldn't be something we will need to worry about too often in the future.

tools like replit, lovable, and v0 are early evidence that people with no coding background can make their own tools. its like when i saw one of my completely non-technical friends make their own personal todo app. or the emergence of all these new founders building mvps in weeks without much coding experience. even for me, i am able to make web applications with decent looking UI despite being more of a backend engineer. im even teaching some of my non-software background friends how to use claude code to make her own prototypes.

with the latest models like Opus 4.5, you have Anthropic engineers who are running multiple agents in parallel, automatically making PRs, and merging them without much oversight. they are providing the right context and guardrails necessary so that the agent does not go rogue and push poor quality code.

software development is converging to a point where not much technical knowledge will actually be needed. i am not even sure if you will need a technical person sitting behind the wheel driving the agent because these models have become more competent from a technical standpoint. this is basically supporting the whole meme of 'this is the era of the idea guy', because now instead of technical aptitude, you need to be able to think of things that the LLM can't. the LLM has totally compressed the executional period of previously needing lots of funding for those 5 engineers to make the 6 month production-ready version one product.

and so what will software development look like in 5 to 10 years? as models get even better, and we get better at managing context (i.e., handling bigger more complex codebases, etc), i think headcount of most of the junior engineers will be significantly reduced because you no longer need headcount to do the grunt work. more notably though, we will see a rise of traditionally "non-technical" users building their own custom "technical workflows". i think this extends beyond the consumer examples i am giving, and means that even businesses will be able to rapidly prototype and develop internal tools to improve processes.

the english language has become another abstraction layer of knowing how to code. instead of needing to know Rust syntax, i need to understand how systems interconnect and have product intuition in order to better deliver an idea. most of all, it is crucial to be able to communicate all this in simply natural language.

i think the ability to just build something is no longer an impressive feat. in other words, just being a generic software engineer with a basic understanding of frameworks like React or libraries like Pandas is not enough. in fact, this is even below the minimum requirement.