The Builder's Paradox: a manifesto for developers

What’s crazy is that us developers are working incredibly hard to build the automation that takes our jobs. We are actively building the system to replace us. With unit tests. And great documentation.

Imagine explaining this to a factory worker in 1985. “So you build the robot, train the robot, review the robot’s work, then the robot gets your desk?” Yes. And we do it with a standup every morning.

I’m not writing this to stop anyone. I build this stuff every day and I genuinely love it. I’m writing it so we do it with our eyes open. So here it is.

Human in the loop. Then human out of the loop.

Right now the industry’s favourite comfort phrase is “human in the loop”. The agent writes the code, the human reviews it. The agent opens the pull request, the human approves it. Everyone relaxes. There’s a human. In the loop. Lovely.

But watch what the loop actually does. Every review we do, every correction we make, every “no, not like that” we type, becomes training data. We are the quality gate the model learns to pass. The better we are at our job in the loop, the faster the loop stops needing us.

First we write the code and the AI suggests. Then the AI writes the code and we review. Then the AI reviews the AI, and we “sign off”. You can see where this is going. Nobody announces the last step. One day the sign off is just a checkbox, and then it’s a cron job.

I’ve worked in a fully agentic team. The human in the loop was very real, right up until the restructure. So if you’re the human in the loop, make sure you’re the one who designs the loop, not just the one standing in it.

The receipts

Let’s be logical about this rather than dramatic. Here’s what the numbers actually say, and what they don’t.

So the total isn’t exploding. 2023 was worse. What’s changed is the reason. In 2023 companies cut because they’d over hired. In 2026, nearly three quarters of the cuts have AI written on them. Same redundancy letter, new excuse. Or new strategy, depending on whether you’re reading it or writing it.

Put those side by side and the shape is hard to ignore. The more of our work we route through AI, the more our work looks like something AI can do, and the easier it becomes for someone with a spreadsheet to write “efficiencies” next to our names.

To be fair, a correlation is not a confession. Interest rates, over hiring and plain old cost cutting are all in there too. But follow the logic. Anthropic’s own data says software development is the single biggest use of Claude, over a third of all activity, and 79% of Claude Code conversations are pure automation, the AI doing the task rather than helping with it. Every one of those is a developer showing a model exactly how their job is done.

And the first people it lands on are the ones using it most. Early career developers have the highest daily AI use, and Stanford’s payroll study found employment for 22 to 25 year olds in the most AI exposed jobs, software development included, fell by 13 to 16% relative to everyone else since late 2022. Older developers in the same jobs? Up. The juniors trained the model, and the model took the junior roles. Efficient, really.

The tipping point

So here’s the obvious question nobody in a board meeting wants to ask. If AI keeps taking jobs, at what point does a society stop working?

There’s no magic number in the research. But history has run this experiment a few times, and the results are pretty consistent.

Here’s the pattern. Societies don’t collapse at 25%. Nobody switches the lights off. What collapses is the politics. Once around a quarter of adults, and half of young people, are out of work for years, people stop believing the system works for them, and they vote, march or riot for something else. Sometimes that’s a New Deal. Sometimes it’s much, much worse.

The research backs up both halves of that. Studies across developing countries link high youth unemployment to political instability. And a study of Weimar Germany found the Nazi vote rose fastest where the government responded with austerity, not just where unemployment was highest. The jobs going is the trigger. What governments do next decides how bad it gets.

Now put AI into it. The IMF estimates around 60% of jobs in advanced economies are exposed to AI, and that roughly half of those could see demand for workers fall. That’s about 30% of jobs, past the line where the Great Depression, Spain and Greece all broke.

In the UK, every 1% of unemployment is roughly 360,000 people. The 25% line is about 9 million people, more than the whole of Scotland and Wales put together. And income tax and National Insurance make up over half of what HMRC collects. A state funded by taxing wages, with fewer wages, has a problem nobody has a slide deck for yet.

Sources for this section: IMF on AI exposure, youth unemployment and instability, AfDB, austerity and the rise of the Nazis, NBER, UK tax receipts.

Who’s next: a timeline

This part is my opinion, not data. It’s based on where AI is already doing the work end to end, how much of each job is text, code or screens, and how little anyone needs a body in the room. Starting, naturally, with us.

1. 2023 to nowHappening now

   Software engineering

   Already here. Junior roles first, then whole teams restructured around agents. Ticket in, pull request out, human signs off. Until the human is a line item someone wants to remove.

2. 2024 to nowHappening now

   Customer support

   Chatbots used to be a joke. Now they resolve the ticket, issue the refund and apologise more sincerely than any of us ever did.

3. 2025 to 2027Happening now

   Copywriting, content and translation

   If the output is words on a screen, it’s already cheaper to generate than to commission. The rates have collapsed before the jobs have, which is how it starts.

4. 2026 to 2028Next

   QA, testing and data roles

   Agents already write the tests and run the suites. Data entry, reporting and dashboards follow, because a model never complains about a CSV.

5. 2026 to 2028Next

   Recruitment and admin

   AI already filters the CVs that AI wrote. Scheduling, inboxes, minutes and follow ups are next, because nobody went into admin for the love of a calendar invite.

6. 2027 to 2029Next

   Accountancy and audit

   Bookkeeping, payroll, tax returns, reconciliations and month end close are rules, numbers and documents. Models love all three. The partner who signs it off stays a while. The juniors who used to prepare it, less so.

7. 2027 to 2029Next

   Legal

   Contract review, discovery and first drafts are already being done in minutes. Paralegals and trainees feel it first. Barristers are safe until judges start accepting a well formatted prompt.

8. 2027 to 2030Next

   Research and analysis

   Literature reviews, market research, financial analysis and the forty page report nobody reads. Deep research agents do in an afternoon what a junior analyst did in a month, and they cite their sources. Mostly.

9. 2028 to 2030Next

   Design, video and marketing

   Not the best designers, the volume work. Banner ads, product shots, social clips, the stuff that pays most people’s bills.

10. 2028 to 2032Later

    Medical diagnostics and admin

    Radiology, pathology and dermatology are pattern matching on images, and models are already very good at it. Medical coding, notes and referrals go first. Regulation slows it down, which for once is the good news.

11. 2030 and beyondLater

    Middle management

    Once the team is mostly agents, you don’t need as many people managing the people. Fewer standups. Every cloud.

12. LastLater

    Trades, care and anything physical

    Plumbers, electricians, nurses, carers, the surgeon actually holding the scalpel. Robots are coming, but slower, and a boiler doesn’t care how good your prompt is.

Meanwhile, in our repos

If you want to see the paradox in its natural habitat, look at the files we now write every week. We used to write onboarding docs for new hires. Now we write them for the thing that replaces new hires.

---
name: senior-engineer
description: Everything I learned in 15 years, in a format a model can read in 2 seconds.
---

# Senior Engineer

## Before changing anything
- Trace the code path end to end. Yes, all of it.
- Reuse the existing patterns. The last developer had reasons.
  Possibly they were good ones.

## When fixing a bug
- Write a failing test first.
- Fix the bug, not the twelve things near the bug.

## When unsure
- Ask the human. While we still have one.

# AGENTS.md

You are working in a production codebase maintained by humans.
Please be kind to them. They are going through a lot.

## Rules
- Keep pull requests small. Reviewers have feelings and limited coffee.
- Do not invent a fix for something that is not broken.
- Do not delete the tests to make the tests pass.
- Run the E2E suite before saying "done". "Done" means it works.

## Architecture
- Follow the boundaries in /docs/architecture.md.
- A developer spent years learning why they exist.
  You are getting it for free. You're welcome.

Signed, a developer who is still here. For now.