Software ate the world. Now AI is eating software one file at a time. If you write code for a living, you have probably felt the ground shift beneath your keyboard. One developer felt it twice in a single year. First, the machine took the writing. Then it took the checking. He still has a job, but only because he changed it beyond recognition.
This is not a doomsday prediction. It is a field report from the middle of the transition.
The First Shift: From Typing to Proofreading
There was a time when his craft was simple to describe: think, type, test, ship. He wrote every function by hand. He named variables, nested loops, and caught edge cases through effort and error. Reading the code was not a separate task. It was creation slowed down and inspected. Quality control and production happened in the same motion.
Then the autocomplete stopped being cute.
AI coding assistants stopped suggesting the next keyword. They started spitting out entire classes, database migrations, and API wrappers. The developer’s hands moved from the arrow keys to the review pane. He lost the act of writing, but he told himself he was safe. He still read every line. He still traced the shape of the logic. He was the editor, the gatekeeper, the human in the loop.
This shift feels manageable because the illusion of control remains intact. You are still touching the code. You are still hunting for off-by-one errors. But your role has already changed. You are no longer the author. You are the first line of defense against the machine’s hallucinations.
The Second Shift: From Reading to Trusting
The second change hit harder.
AI agents grew fast enough to work for hours without human oxygen. They could spin up scaffolding, refactor legacy modules, and patch dependencies while he stepped away for coffee. The throughput was no longer the problem. His own reading speed was.
To keep the project moving, he had to stop reading every line. There was no other option. The factory could produce faster than he could approve. Trust became a requirement, not a virtue. He went from proofreader to operations manager, overseeing a system that generated more code in a sprint than he used to write in a year.
He described the feeling honestly: he was no longer a coder. He was managing a machine factory where the conveyor belts moved faster than his eyes.
The Bottleneck Moves Uphill
Here is the part that is easy to miss. The bottleneck did not vanish. It relocated.
When humans wrote code, the delay was between the brain and the IDE. The developer was the slowest moving part, and that was acceptable. Once AI took over authoring, the choke point became human review. Then, once agents could run in multi-hour loops, the choke point became human decision-making. What should the agents work on? What is safe to automate? What requires a second look?
The factory produces faster than the manager can approve. The problem is no longer the machine’s speed. It is the human’s capacity to absorb, judge, and direct.
If you ignore this migration of the bottleneck, you become the obstacle that the next tool is built to remove.
Three Rebuilds, One Career
Surviving both waves required rebuilding his role three separate times.
First, he moved from writing code to reading code. That meant learning to skim with skepticism, hunting for subtle bugs in boilerplate generated by a model trained on open-source repos and forum threads. It meant developing a nose for when a suggestion smelled right but was subtly wrong, a skill entirely different from authoring from scratch.
Next, he moved from reading code to building workflows. This is where many developers are landing right now. You are not fixing functions; you are designing the environment where functions get fixed. You write prompts, craft system instructions, set guardrails, and configure CI gates. You build the plumbing that connects the agent to the compiler, the linter, and the test suite. Your output is less syntax and more orchestration.
Finally, he moved from building workflows to deciding what to build. This is the thin air at the top. You are choosing the architecture before the first file is generated. You are defining the constraints, the interfaces, and the business logic that the AI will fill in. You have become a product thinker who happens to speak in systems.
Si se hubiera quedado en la etapa uno, lo habrían reemplazado por completo. La única razón por la que todavía cobra un sueldo es que ascendió a la etapa tres.
Pero el ascenso es