← Back to Blog

Three Years That Changed Software Development

For the first time in my career I'm watching experienced developers struggle, not with the technology, but with the pace of it. Keeping up has become a second job, and software is just the first industry to feel it.

By John Croucher
AISoftware DevelopmentOpinionCareers

I have been writing software for over 20 years, and technology changing underneath you has always been part of the job.

Languages come and go. Frameworks get replaced. Servers moved into the cloud. We went from desktop applications to websites, then mobile apps, containers, serverless, DevOps, microservices and whatever else somebody decided we needed that week. There has long been a joke that a new JavaScript library comes out every week.

You learn it, adjust, and keep going. That has always been the deal.

The last three years feel different.

It isn’t just another technology to learn. The tools are changing what the job actually is, and they are improving fast enough that something you learned six months ago can already feel outdated.

Changing code editors used to be a once every five to ten years event. Now there is a different way of working every few months. And past changes usually touched one part of the work: one language, one framework, one editor. This one touches everything. The job I do today is not the job I did 12 months ago, and that is true for everyone, including the people who resisted it.

For the first time in my career, I am seeing experienced people genuinely struggle with the pace of change. Not because they aren’t good developers. Because keeping up has effectively become another job.

It Started as a Better Autocomplete

When generative AI first turned up in software development, it was easy to understand. You started writing some code and it suggested the next few lines. Useful, sometimes impressive, sometimes completely wrong.

Then we started asking ChatGPT questions instead of searching Google or Stack Overflow.

Then it could generate functions. Then entire files. Then you could point it at an existing codebase and ask how the system worked. Then it could change multiple files at once. Then it could run the application, see that its change failed, read the error, modify the code and try again.

Now we have coding agents that can be handed a task and left to work through it almost like another developer.

The move from “finish this line of code” to “here’s the problem, go work on it” happened remarkably quickly, and it changes what a software developer does.

Saving Time Has Not Meant Having More Time

There is a funny thing happening with AI productivity. It absolutely can save time. That does not mean developers suddenly have spare time.

A 2026 BairesDev survey of 705 developers found they estimated AI was saving them an average of 13 hours of coding time per week, almost double the seven hours they reported a year earlier.

Those hours were not disappearing from the working week. Developers reported spending more time reviewing AI-generated code, debugging problems the AI introduced, and learning the constantly changing set of tools around it.

13 hrs Coding time saved per week Up from 7 a year earlier
9 hrs Spent learning per week Up from 4 a year earlier

BairesDev 2026 survey of 705 developers.

That matches what I see. AI reduces the amount of typing. It doesn’t necessarily reduce the amount of thinking, and in some cases it increases it.

Instead of writing 100 lines of code yourself, you might have 500 lines generated in a few seconds. You still need to work out whether those 500 lines are correct, and unlike code you wrote yourself, you don’t have the mental model that came from building it.

So the job moves upward. Less typing. More reviewing, more architecture, more checking assumptions, more understanding how different systems fit together. More deciding whether something the computer confidently produced is actually a terrible idea.

I wrote recently about shipping code you don’t understand. This is the other side of that problem. The amount of work we can generate is growing faster than the amount of work humans can properly understand.

Everyone Is Using It and Nobody Completely Trusts It

That contradiction shows up clearly in the numbers. Stack Overflow’s 2025 developer survey had more than 49,000 respondents. 84% said they were using or planned to use AI tools in their development work. At the same time, 46% said they actively distrusted the accuracy of its output. Only 33% said they trusted it.

So the industry is rapidly adopting a tool that much of the industry doesn’t trust.

That sounds ridiculous until you use it. I use AI every day. I also regularly watch it produce something completely wrong with absolute confidence. Those two statements are not contradictory. The productivity benefit is large enough that not using it increasingly feels like a disadvantage.

That creates another pressure. You don’t just need to know how to build software anymore. You need to know which models are good at which jobs, which coding agent works best, how to give it context, when to let it work on its own and when to stop it, how to review what it produced, how to structure a project so an AI can navigate it, and how to keep it away from things it shouldn’t touch.

By the time you are comfortable with that workflow, a new model comes out and changes the equation again.

The Exhausting Part Is the Acceleration

There is always hype around new technology, so I am usually cautious about extrapolating too far.

One piece of research I keep coming back to is from METR, an organisation that measures how well AI systems complete tasks autonomously. Rather than asking whether a model can pass some artificial benchmark, they ask something much easier to understand: how long would the same task take a competent human?

Their research, mostly covering software engineering, machine learning and related work, found that the length of task frontier AI systems can successfully complete has been roughly doubling every seven months.

Not performance improving by a few percent. The length of the task it can handle, doubling.

That doesn’t mean an AI can replace somebody doing a week-long project. Real jobs have context, ambiguity, meetings, politics, customers and all the messy things benchmarks don’t capture, and METR is clear about those limitations themselves.

The trend still matters, because a seven-month doubling time is extremely hard for humans to adapt to.

Imagine working with a new employee who doubles in capability every seven months. You finally work out what to delegate to them and what they’re bad at. Seven months later, that understanding is wrong. Then it happens again.

That is roughly what working with these tools feels like right now.

A Couple of Months of My Dad’s Work, in Three Minutes

My dad used to work as a purchasing officer, and we were talking about AI and the kind of work he used to do.

He mentioned a project from years ago where he needed a particular product manufactured. He had to work out what it should look like, its dimensions, how it could be built and what components it needed, document all of that properly, then find companies capable of making it. The process took him a couple of months.

While we were sitting there, I described the requirements to an AI. A couple of minutes later it had produced a product concept, a mock-up, technical drawings, a parts list, suggested materials and the types of manufacturers to approach.

He just sat there looking at it.

Two months of his work had been compressed into about three minutes.

That doesn’t mean the project was finished. Someone still needs to verify the measurements, check the engineering and validate the materials. Manufacturers need contacting, quotes need negotiating, a prototype needs building and the finished product needs testing. There is a huge difference between an AI producing a plausible design and thousands of safe, working products coming off a manufacturing line.

But that wasn’t the point. A significant chunk of a process that once took weeks of research, drafting and coordination had collapsed. Even if the AI only gets you 70% of the way there, that is an enormous change.

Software Is Just Experiencing This First

Software development is particularly exposed because almost everything we work with is already digital. The code is text. The documentation is text. The error messages are text. The tests run automatically, and the development environment can be controlled by software.

An AI can write some code, run it, inspect what happened and try again without ever leaving the computer. That makes software an almost perfect environment for this technology.

Other industries have much more friction. You can generate a CAD design instantly, but someone still has to manufacture the part. You can generate a legal document instantly, but somebody qualified still needs to check it. You can produce a medical analysis instantly, but the consequences are considerably higher than a failed unit test. You can design a building instantly, but there are still engineering standards, council approvals, construction, and a very physical building at the end of it.

That friction will slow things down. It doesn’t mean those industries won’t change. It means the change will look different.

The Job Doesn’t Disappear, the Valuable Part Moves

This is where a lot of the conversation around AI gets too simplistic. People ask whether AI will replace software developers. That is probably the wrong question.

Computers have been replacing individual tasks inside jobs for decades without removing the whole job. What tends to happen is that the valuable part of the job moves.

Developers used to spend far more time on memory management, configuring servers and manually deploying applications. Tools automated much of that. We didn’t stop needing developers. We built more software.

AI may do the same thing on a much larger scale. If writing routine code becomes cheap, understanding what should be built becomes more valuable. So do architecture, domain knowledge, knowing whether an answer makes sense, being able to talk to customers, and taking responsibility for the result.

That last one is important. The AI doesn’t get called at 2am when production goes down. Someone does.

The Learning Pressure Is Leaving Tech

There are signs the pressure developers feel is starting to spread.

PwC’s 2026 Global Workforce Hopes and Fears survey covered 49,364 workers across 48 countries and regions. Nearly two-thirds said they had needed to learn new tools or technology for their job over the previous year. Only 51% said they had access to the learning and development resources they needed, down from 59% the year before. PwC also found a large group of workers who felt the change was leaving them behind rather than benefiting them.

That should concern businesses more than whether somebody can write the perfect prompt.

You cannot keep increasing the rate of change and assume employees will absorb it for free. Learning takes time. Experimenting takes time. Keeping up takes time. If every productivity gain just means people are expected to produce more while also learning the next tool, eventually that gain starts to look a lot like burnout.

Software development may just be getting there first.

Three Years Is the Part I Can’t Get My Head Around

This hasn’t played out over a generation. It has happened in roughly three years.

Three years ago, most developers wrote the overwhelming majority of their own code. Today I can describe a feature and have an AI inspect an unfamiliar project, change multiple files, write tests, run them, find its own errors and keep going.

It still gets things wrong. A lot. There are plenty of problems it cannot solve, and tasks where I spend more time arguing with it than I would have spent doing the work myself.

Focusing on those failures misses what is happening. The important thing isn’t where the technology is today. It’s the slope of the line.

Every few months the boundary moves. Things I wouldn’t have trusted an AI with last year, I now give it routinely. Things that were impressive demos two years ago are boring features built into everyday tools.

Developers Are the Canary in the Coal Mine

I don’t think software developers are uniquely doomed. If anything, we are unusually well positioned. People in technology have spent their careers learning new tools. We are used to our knowledge going obsolete and expect to retrain constantly.

And even we are finding the current pace uncomfortable.

That is why what is happening in software is worth watching if you work somewhere completely unrelated to technology. Your job probably won’t vanish because somebody installed an AI tool. The more realistic change is subtler.

A task that takes a week becomes a day. Something that needs five people needs two. Research that took a month is done before lunch. The first draft becomes effectively free, and checking whether it is actually correct becomes the important part. Then six months later another part of the process moves.

That is the change businesses and workers need to prepare for. Not a science-fiction morning where everybody wakes up to find AI has taken their job, but a thousand smaller changes where the boundary between what humans do and what computers do keeps moving, faster each time.

Software development is already living through it. The rest of the world may just be a few steps behind.