It is 6am, at least somewhere in the world. I am in a transatlantic flight on and I refuse to count the number of miles. I am a “road warrior” as they say. Spend two to three weeks per month traveling. And that affects my health specially with the food I eat. Like many people, I decided to start taking that seriously and went to hotel gyms and started looking at my diet. Specifically, counting calories and macros. Now, that can work great when you are home, but when you travel so much it is very difficult to “count calories” or really know what you are taking in. The usual tools don’t cut it. So I decided to fix it. What if I could send a whatsapp picture to a tool and automatically have all the information gathered, macro nutrients, energy, etc. And recommendations for my diet done? So, as I was around Greenland I started vibe coding my new app. Created a whatsapp number, built the whole app and connected to some nice LLMs, tested it on the plane with the food they served me… et voila! I just had in a matter of 3-4 hours a full application that served my purpose (and still does). For almost no money and most importantly. Without seeing a single line of code. Only writing in english what I wanted.
Should I claim this “app” is a business? Of course not. Probably it only serves me and my need today. But again: I built it, to my needs and taste, in a few hours and with very minimal money investment (maybe 10$ in tokens and opening an account on Twilio). And I created more value to myself than many other apps I should’ve gone out and research and evaluate.
We are in that point where careers are redefined. Where some people will lose their jobs and some will end up building multi billion dollar companies. The whole paradigm is changing completely. Some are realizing that but some others are still thinking they are safe.
But let me start from the beginning.
Software Engineering was always considered a craft. A type of art that helped people create new things. Worlds, systems, operations, applications. You name it. We were creators. And with that came a sense of pride. We loved what we typed. It’s poetic structure. Its maintainability. We saw ourselves as smart mammals capable of building mutable machines that replicated the world and were able to mutate into complex systems that solved the world problems. From trading algorithms, to dam-flood-control systems, all the way to dumb mindless mobile app games.
We have been always looking for the best and most efficient way to solve a problem. Sometimes holistically, sometimes within a specific realm. We have added layers and layers of abstraction between us and what runs our programs and expanded our reach to the whole internet. From managing the inner layers of the memory and our CPU cache, to super high level javascript abstractions where we don’t even need to think how to trim a string.
As new companies were built and digital was the new natural, we were sought after like scientists in the era of the Manhattan project. Companies paid more and more for software developers. University degrees did not matter that much anymore. Bootcamps started popping up everywhere so that people could reinvent their careers and join the gold digger fever. Parents wanted their 6 year olds to speak Python before english. Thousands of people changed careers to be part of the new cool. Software development was the quickest way to make a six figure salary. Move away Wall Street. Here come the coders.
Over time we have matured the industry. We have created career ladders, organizational frameworks, quality systems, design processes. And whole charts that make a “path” for any new entrant to know what’s ahead. From junior engineers, to staff, to CTOs and VPs of Engineering. Beautiful stuff. Now we know how to build stuff, how to organize ourselves and how much it will cost (well, not quite. It is still a very immature business, but you get the idea).
And suddenly everything changes. The rules of the game are completely thrown away. Suddenly your QA processes are completely automated. Your frontend development ends up being a commodity and you require a fraction of the man power you needed back then (ironically, we are back to the “surgical team” concept by “The Mythical Man Month”). I am not talking about “vibe coding” a prototype like my experience at the beginning of the article. But about the whole rearchitecture of the Software Development Lifecycle.
In the past (present) we built software in a variety of ways. But they all boiled down into the same concept. There is an idea or request, a design phase, a development phase, testing or certification, deployment and monitoring and feedback. Depending on your flavour the phases can change order or intersect each other. Testing may precede development just to take over right after. You may deploy continously, or do the process in very small chunks for quick iterations and feedback loops. OR you can just shoot from the hip and deploy what a product manager fancies whenever it is safe for customers. But from a high level it all does the same thing. Code, Build, Deploy, Repeat.
And because our platforms have become so complicated, we have especialized ourselves in ways that we could not even think of ten years ago. A fullstack developer is now regarded as a shallow generalist. Frontend developers focus in unique frameworks following their own cargo cults. Data has its own universe and don’t get me started with mobile development. Specialization, like anywhere else, brought depth, capabilities and access. But value only came in certain cases. Many teams ended up building technology for the sake of technology instead of customer driven value.
And now, at the dawn of Agentic AI we see that the two three most important and ROI generating use cases are around streamlining (and replacing) customer support flows, QA in software development and the muscle of building the actual software. The princesses of the industry, the high salary individuals that were sought after like financial quants in the late 90’s… are now being automated.
I am not talking just about using Github Copilot or other tools to be faster and more accurate. I am talking about a full replacement of the engineering pipeline. Imagine a system that:
- You write a detailed flow on how you want your application to behave (say, a web application). You describe it in good detail (with your voice). Flows, interactions, users, etc. An LLM gives you a near perfect Product Requirements Document that is better and more complete than what any product manager has ever created. Of course you can iterate, add and change things, give more context and examples, etc. All until you get your desired outcome.
- You give that full PRD to your vibe coding system of choice. Now there is enough players in the market to get them to fight for new and better features and models continuously. It is not about one of them being the best today. It is about keep the race on and cost low. With that you achieve a playable and deployable prototype (or even a hint of an MVP) that you can show to people. Investors, clients, or your grandmother. This is your first feedback loop. You are recording those sessions of people using it and giving the feedback to your system in point one. With that feedback it will refine the PRD and then you can build the v2 of your prototype. Rinse and repeat until happy.
- Now you have a full Product Definition Document, user stories, acceptance criteria, working prototypes (instead of just a mockup). So you can create that document in a Markdown document and put it in an empty Github repository. The party starts.
- A project manager agent will divide the work in small chunks and will order it properly with dependencies. Then it will create automatically Github issues (if you prefer JIRA that’s on you). With unique and specific definition and description of what is intended.
- Whenever a ticket is created, an army of “AI Coders”, that are just waiting, start picking up tasks. You can parallelize them if you want. They will write the tests, build the code, make the build, fail, iterate. Test Driven Development becomes the norm now. When a task is completed as per the acceptance criteria, they send their code to the repository in a Pull Request.
- Another agent reviews the pull request. It may have flaws, it may not run properly integration or end to end test. If so, it send it back to the agent developer. Keep iterating until it can be merged.
- The last agent takes the latest tag with all the PR’s integrated, makes the build and deploys it into production. YOLO. A monitoring agent that’s hooked to your profiling and APM solution of choice detects any error immediately by reading the logs before a human has even looked at his pagerduty. Creates an urgent ticket, a developer agent fixes it, PR, deploy, repeat.
That cycle typically lasts weeks for many companies. Months for large enterprises. Here, this is done in minutes or hours. Continously, 24/7, 365 days per year.
We are effectively redefining everything. The concept of technical debt, of migrations and upgrades. Of on-call rotations. Even of Agile and bloody SCRUM (about time we start opening that can of worms). 3 weeks to provide value units to customers? I want 3 hours! All the roles in the software world are changing in front of our eyes. The technical lead, the testers, backend, frontend, DevOps…. The agents learn from us, connect from us and react better and faster than us. We just define what we think our customers want.
Am I being optimistic? perhaps, but I am seeing those systems already working at a small scale. Startups that would have required a quarter of a million to start moving, now they just need a savvy builder and a couple billions of tokens (and a fraction of the time). And this is just the beginning. ChatGPT came out end of 2022. 2023 saw the advent of LLMs everywhere and in 2024 many companies went into production with their use cases. 2025 and beyond is where Agentic AI is taking over. New and better models come every month, improved frameworks, etc. Things can only improve. How are you going to adapt to this new reality?
Software engineering is dead, long live software engineering.