We are not watching the end of software engineering. We are watching it get split in half.

Nobody is getting mass fired yet. Almost everyone's job has already changed. All the hype about replacement and all the counter-hype about nothing changing are both wrong. What is actually happening is quieter, messier, and far more permanent.

This is not a replacement event

"The dumbest thing I've ever heard" was AWS CEO Adam Selipsky's reaction last year to companies laying off junior engineers and replacing them with AI. He was right, but not for the reason most people quoted. It is not that AI can't do junior work. It is that AI does exactly the opposite: it eats the work everyone was already bored of doing first.

This is not the robot apocalypse narrative. There will be no morning where an email goes out saying everyone is replaced. There will just be a Tuesday where you realise no one has written boilerplate CRUD endpoints by hand in 18 months.

54% of all code committed today is AI generated. That number is from this month. It was 28% one year ago. No one stopped the world to announce this. It just happened.

The 80/20 split is real

Micky Arun runs an FCA authorised payment platform. They have run agentic coding tools for over a year.

80% of their codebase is now written almost entirely by AI. API scaffolding, Zod schemas, docker configs, Traefik labels, env var migrations across 12 repos. Work that used to take junior engineers half days or entire weeks now completes in minutes.

No one on the team misses this work. No one is proud of this work. No one learned anything doing this work. It was just the tax you paid to get to the interesting part.

The other 20%? The agent will write it. It will pass all tests. It will look correct. And it will silently break payment settlement at 2am on a Saturday.

This is the split no vendor will tell you. AI does not sort work by seniority. It sorts work by consequence. Low risk, repeatable, pattern matching work vanishes first. High consequence, high ambiguity work stays. And the line between them is almost never visible until something breaks.

The cost wall no one warned you about

This is the most underreported development of 2026.

Microsoft cancelled almost all internal Claude Code licenses last month. Every team working on Windows, Surface, Teams and Outlook will migrate to Copilot CLI by June 30. Microsoft is one of Anthropic's largest investors. They own Azure. They still could not afford the bill.

Uber burned through their entire 2026 AI coding budget in the first four months of the year. Their CTO told internal teams they were going back to the drawing board. They had run leaderboards encouraging engineers to use the tools. Usage exploded. Costs went vertical.

This is not a teething problem. The entire business model for agentic coding tools is per token. Every extra loop the agent runs, every re-read of your codebase, every redundant tool call is revenue for the provider. Productivity and cost are perfectly correlated. The better the tool works, the faster your bill runs away.

No one has solved this. Not the biggest companies on the planet. Not the people building the models.

Good enough is the new default

AI does not write bad code. It writes code that is exactly good enough.

It will pass every test. It will handle every edge case you mentioned. It will ship. It will never be elegant. It will never have good variable names. It will never have a comment explaining why a decision was made, only what was done. It will be nested one level deeper than it needed to be. It will make the next change 10% harder.

No individual change will kill you. 1000 of them will. Over two years you will end up with a codebase that works perfectly, that no human understands, that no human can modify.

This is the quiet technical debt no static analyzer will catch. It is not bugs. It is lack of care. AI optimises for closing the ticket. It never optimises for the person reading the code six months from now.

The hidden damage: collective thinking

All of the above is individual. The worst effects are showing up at team level.

Individual productivity is up almost everywhere. Team alignment is down.

Engineering was never an individual activity. It is a coordination game that runs on shared mental models. Before AI, when you built a feature you worked through the hard parts together. Everyone saw the tradeoffs. Everyone understood why decisions were made.

Now one engineer can generate an entire service in an afternoon. No one else saw the choices. No one else argued about the tradeoffs. No one else built the mental model. The code ships. The team drifts.

Teams that adopted AI without changing anything else are now faster, and much dumber. They ship more features. They understand less of the system they are running. And they will not realise this until the first major outage.

The job is being redefined, not deleted

Go back to the translator parallel. AI did not eliminate translators. It turned them into editors. 80% of the typing work is gone. The remaining 20% is the hard part: knowing when the machine got it wrong, fixing the idioms that do not translate, understanding the audience.

That is exactly what is happening to coding.

You will not spend most of your day typing code. You will spend most of your day reading code that was generated for you. You will spend your day judging. You will spend your day understanding consequence. You will spend your day being the person that knows why.

This is not better or worse. It is just different. It is also much harder. It demands much better judgement, much deeper system understanding, and much more experience. The bar for being an effective engineer just went up, not down.

There is one final cost that no one is measuring. It is the slow erosion of actual conversation. When you ask a question, and get back a screenshot of ChatGPT. When you open a GitHub thread and half the comments are just pasted AI output. We built tools to make us faster. We accidentally built tools that let us stop talking to each other. That might be the most expensive trade of all.

What works right now

Teams that are navigating this well are not banning AI. They are not mandating it. They are doing three boring, unremarkable things that will never appear on a conference keynote.

First, they draw explicit lines. AI writes boilerplate, tests, migrations and refactors. Humans write all code that touches state, payments, permissions or error handling. No exceptions. No arguments.

Second, they require human reasoning to be written down. AI output does not count as documentation. Every change still requires a human to write one paragraph explaining why it was done that way. Not what was done. Why.

Third, they stopped measuring individual output. Speed is no longer a useful metric. Understanding is.

None of this is exciting. None of this will make you go viral on LinkedIn. This is just what normal adaptation looks like.

This transition is not over. It has barely started. But we already know enough to say this much: AI will not end software engineering. It will just separate the people who understand what they are building from the people who do not. That was always going to happen eventually. AI just moved up the deadline.