Back to all posts
·6 min read·Like or comment

AI Has Reached an Inflection Point in Software Development

Get new posts by email. No spam, just an email when I publish something new.

TL;DR

  • AI coding tools have gone from autocomplete to autonomous agents that run for days and produce production-ready code. After 25 years, I don't expect to write code by hand again.
  • The latest models crossed a quality threshold where line-by-line human review is no longer the bottleneck, making agents roughly 100x more productive.
  • It takes engineering skill to set up: agents need guardrails and feedback loops (types, tests, a browser) so they can check their own work.
  • Companies should push agents into their process and budget generously for tokens. Developers should aim to be the best at using AI.
  • Software is just first. Every job done on a computer is next, and I expect short-term pain but more prosperity in the end.

Something shifted in the last couple of months. AI coding tools went from "useful autocomplete" to autonomous systems that can run for days. The gap between what's possible with and without AI is now so wide that ignoring it means falling behind.

To put it plainly, after 25 years of writing code, I don't see myself writing a single line of code from now on.

I've been anti-hype on this until now because it wasn't quite there. That's changed. I'm talking about building real software day to day in 2026, and demos that fall apart in practice don't count. The tools are now good enough to change how much software costs to build, how fast it gets built and who can build it.

From autocomplete to agents

I've been using AI as a coding partner since 2024. For most of that time it was back and forth: checking every line, and often turning it off to correct mistakes.

In the last month I stopped reviewing. I let it run by itself, sometimes for days, and what comes out is production-ready code.

This won't be everyone's experience. You need to know how to set it up and build the right guardrails. If you know what you're doing, that takes a couple of hours. After that, you'll never write code by hand again.

What changed?

The latest models, Claude Opus 4.5 and ChatGPT 5.2, crossed a threshold. They're now better than humans at most coding tasks.

When AI was about 70% as good as a senior engineer, human review was the bottleneck. You couldn't leave it unsupervised, and every line needed checking before it went to production.

Now it's around 120% as good. That doesn't sound like a huge jump, but it changes the workflow completely. If you don't need to review code line by line, humans stop being the bottleneck. And while it's only a bit better on quality, it's about 100x faster.

So crossing that line makes coding agents about 100x more productive in practice.

Autonomous agents

Now that you can rely on these tools, the bottleneck is how many features you can spec out.

You can get past that too: set high-level objectives and let the AI come up with its own features.

A planner invents a feature, then breaks it down into tasks and subtasks. A coding agent picks up one task at a time and does it. Now you have a system that keeps building product features it invented by itself.

From a single one-page plan, I built contrary.markets in 2 days, with no guidance beyond that plan. It invented features like leaderboards, achievements and paper trading (trading with pretend money) by itself, then scoped and built them.

Setting up your tech stack

This takes engineering skill to set up properly.

You need a stack that works well with guardrails. The agent needs to test its own work: types are correct, tests are comprehensive and database migrations work. It also needs a web browser so it can click through the app and take screenshots for feedback.

Feedback is key. Anything the agent can't check isn't worth doing autonomously. If it writes code to send an email but can't verify the email arrived, it's flying blind. An engineer's job now is to give agents feedback loops so they can correct themselves.

Beyond software

The same ideas apply to any digital problem where there's a clear way to tell if it's been done right: analysing stocks, solving maths problems, automating tasks like booking flights. All of this is coming this year.

Creative work

It doesn't do so well where it can't get clear feedback, or enough of it. It can write a book, but maybe not a good one. It can't tell if what it wrote lands with a human, so it can only follow patterns of what worked before.

Given AI seems to beat human intelligence in many areas, what makes a human able to write a good book? It can't just be structure, because AI knows how to structure a sentence. I think it has to be about pulling life experiences into the writing so it has an emotional impact on the reader.

AI doesn't have these experiences. But it might be able to read about other people's experiences online and use that as a basis for good stories. Time will tell.

For companies

Every company will move at breakneck speed this year, building faster and better than ever before. So don't get left behind.

Set your team the objective of getting AI agents into your process as much as possible. This improves quality and speed at the same time.

Give them the right tools. Don't be cheap: buy good laptops and set a high budget for AI agent tokens. Max out Claude and Cursor subscriptions, and budget a couple thousand extra to go over limits.

For developers

Developers have long been the ones taking other people's jobs. Admin jobs largely disappeared because developers built systems to automate them. Now it's happening to developers, and many refuse to accept it.

A lot of them are resisting it, scared of losing their jobs and their passion for writing code.

Things change. That's life. You can bunker down and hold onto the past, or embrace change and enter the new world.

People who stay optimistic, embrace change and learn new skills will always have a job. People who fight it will become less productive and less hireable.

My advice to developers: become the best at using AI to get the best outcomes, in the most productive way possible. Do this and you'll be the unicorn everyone wants.

Where this goes next

It's taking over software development first, mainly because software developers have the skills to use these agents.

But it will soon come for every job that can be done on a computer. There's nothing special about software that makes it easier to automate. It just happened to be first.

Where does this leave humans?

For now, humans are still needed to make judgement calls. But more of that will be automated away.

Humans will focus more on high-level direction, working with other people and building communities.

I don't know exactly what this means for us. I think there might be short-term pain, but the productivity boost will make us more prosperous in the end.

The "lump of labour fallacy" (the idea that there's only so much work to go around) has been disproven many times. When a new technology comes along, like the sewing machine, there's a dip in jobs at first. But the price of clothes drops, demand goes up, and more jobs get created operating sewing machines than were lost doing it by hand.


I'm thankful for this technology. I've been writing software for 25 years and loved every minute of it. But writing software by hand was becoming mentally taxing and tiring. I've found a new passion for software. I can have an idea, build it and see the results in minutes.

I hope everyone else finds the same joy in this that I have right now.

More to read