This week a Chinese AI lab released a research prototype that watches developer group chats and GitHub repositories on its own initiative, then proposes code fixes without being asked. DeepSeek's Super-Agent is openly a work in progress, but one line in its documentation matters more than the tool itself: much of the engineering was done by coding agents. The people building the replacement are using the replacement to build it.
Around the same time, a joke video called "Software Engineers 2028" circulated among developers, imagining the profession reduced to supervising machines and soothing stakeholders. It landed because it touches a real anxiety. If agents can read the chat, write the patch and open the pull request, what exactly is left of the job?
History offers a surprisingly precise answer, and it begins with the floppy disk. The disk did not vanish because people stopped storing data. It vanished because the function migrated to better carriers, and the object that once defined the function became a museum piece. Job titles follow the same physics. The typist disappeared not when writing ended but when writing detached from a dedicated machine and a dedicated person. Functions migrate; titles linger for a while and then quietly retire.
Software engineering is now mid-migration. The part of the job that is pure transcription, converting a clear intention into syntax, is moving to machines at remarkable speed. GitHub's own research has found developers completing routine tasks dramatically faster with AI assistance, and every major lab now ships an agent that can hold a coding conversation. The typing layer is going the way of the floppy disk.
But a profession is not one layer. Strip the typing away and three harder layers remain. The first is deciding what should exist at all: framing the problem, choosing the trade-offs, knowing which requirement is real and which is a stakeholder's guess. The second is verification: reading machine-written code with the suspicion it deserves, because an agent that writes fluently also fails fluently. The third is ownership. When the system breaks at 2 AM, someone must carry the legal, financial and moral weight of the failure. No licence, contract or court assigns that weight to an agent.
The labour market data tells the same story from the other side. Demand is shifting rather than collapsing. Job postings increasingly ask for engineers who can direct AI tools, review generated code and design systems that include non-human contributors. The World Economic Forum's future-of-jobs analysis consistently places AI and machine learning specialists among the fastest growing roles, even as it expects routine programming tasks to automate away. The pyramid is getting narrower at the base and wider at the top, which is what happens to every profession when its entry-level rung mechanises.
There is a genuine risk in that shape, and it deserves honesty rather than comfort. Senior engineers were once made by doing junior work slowly and badly. If machines absorb the junior work, the apprenticeship pipeline breaks. Companies that celebrate cheaper code today may discover in five years that they have no one capable of reviewing the code their agents write. The rational response is not to ban the tools but to redesign the ladder: pair juniors with agent output deliberately, make review a taught skill, and treat supervising machines as the new craft rather than an admission of decline.
So will the title software engineer exist in 2028? Almost certainly yes, because titles die slowly and institutions are conservative. Hospitals still employ physicians centuries after the role stopped meaning what it meant. But the title will describe a different daily reality: less typing, more deciding; less writing, more reading; less producing, more answering for what was produced. The floppy disk lesson is not that things disappear. It is that the object vanishes while the need survives, and the people who thrive are the ones who follow the need rather than defending the object.
For leaders, the action is concrete. Map which layers of your engineering work are transcription and which are judgment, then hire and train for the judgment layers explicitly. Build review capacity before you need it, because agent-written code still requires a human who understands it. And watch the apprenticeship question in your own organisation before the market asks it of you. The title will take care of itself. The pipeline will not.
Think mathematically. Engineer intelligently. Lead strategically. Build responsibly. Scale globally. Create real business value through artificial intelligence.
