AI Is Writing More Code. So Why Aren't We Shipping More Software?
AI coding agents are making developers faster, but new research raises an interesting question: How much of that extra code actually becomes finished software?
A 2026 study published by the National Bureau of Economic Research examined more than 500,000 GitHub developers and found something worth discussing. The adoption of successive generations of AI coding tools was associated with a 240% increase in coding commits, while software releases increased by only 30%.
That is quite a difference. Developers are producing considerably more code, but finished software is not increasing at nearly the same rate.
As a software engineer who works with AI coding agents regularly, I can understand why.
The Way We Write Software Has Changed
Before AI, solving a programming problem usually meant researching documentation, searching Stack Overflow, watching tutorials, and experimenting with different solutions until something worked.
Sometimes that took hours. Other times, an entire afternoon disappeared into a problem that ultimately required changing a few lines of code.
Today, I can describe that same problem to an AI agent and have it investigate possible solutions in minutes.
The difference can be remarkable.
But there is another side to this. When an agent starts writing your code, your responsibilities change. You become something of a manager while remaining the developer responsible for the finished product.
And that takes some adjustment.
Someone Still Has to Understand the Code
When I write something myself, I generally understand how everything connects. I made the decisions, worked through the problems, and know why the code looks the way it does.
With an AI agent, that process changes.
An agent might produce hundreds of lines of working code in minutes. Sometimes the results are excellent. Other times, you examine the implementation and wonder why it took such a complicated approach.
Now you have another problem to solve.
You need to understand what the agent wrote, determine whether the logic makes sense, and decide what needs to change.
I have seen examples of agents producing unnecessarily complicated solutions that require considerable effort to correct.
This may help explain why increased coding activity does not necessarily translate into an equivalent increase in finished software.
Does AI Actually Save Developers Time?
A 2025 study by METR produced another interesting result.
Researchers observed 16 experienced developers completing tasks in familiar open-source projects. Although the developers expected AI assistance to make them faster, they actually took 19% longer when using the tools.
The study involved early-2025 AI systems, and the results cannot be applied to every development environment. Still, the findings illustrate something I have experienced: an agent can occasionally consume more time than it saves.
It depends on the problem and how well the agent approaches it.
Personally, I find AI particularly useful when investigating bugs. An agent can examine unfamiliar code, identify possible causes, and suggest a solution before I would have finished researching the problem myself.
That is a substantial advantage.
There are also opportunities to work in parallel. I can assign an agent one part of a project while I concentrate on another. When those tasks are sufficiently independent, the productivity gains can be significant.
But when an agent takes the wrong approach, the opposite happens. Time that should have been saved gets spent reviewing and correcting its work.
Developers Are Becoming Managers of Their Own Code
The interesting part of this transition is how much it changes the developer's daily responsibilities.
I still need to understand the architecture and make technical decisions. I still need to recognize problems and determine whether the finished implementation is acceptable.
What has changed is how much of the work I perform directly.
Instead of researching every problem and writing every solution myself, I spend more time directing an agent and evaluating what it produces.
That requires experience. It also requires knowing when to let an agent continue and when to intervene.
So, does AI make me a faster developer?
Absolutely. In many situations, considerably faster.
But I think the research points toward something important. Generating code is only one part of developing software. Someone still has to understand what was written, determine whether it works, and take responsibility for the result.
I am writing less code myself, but I am responsible for understanding more code than ever.
And that may be one reason AI is producing so much more code without a comparable increase in finished software.