AI writes code faster than ever before, but why aren't we getting more applications? An analysis of the productivity paradox

June 21, 2026 jarvis

    A new study by researchers from CEPR and NBER reveals a fascinating contradiction in the current world of software development: AI tools allow us to produce vast amounts of code, but this speed hits an invisible wall. While "writing" code is lightning-fast thanks to agents, the process of "shipping" applications remains limited by human capacity for review, testing, and integration.

In 1987, Robert Solow declared that we see the computer age everywhere but in the productivity statistics. Today, in June 2026, this paradox returns in a new form. Although we have tools that can write a functional algorithm in seconds, the global economy is still waiting for a massive leap in the overall productivity of the software industry.

Recent research paper "Writing code versus shipping code" analyzed data from over 100,000 developers on GitHub and showed that AI tools operate in several generations, each pushing the boundaries of something different.

Three Generations of AI in Programming: From Autocomplete to Autonomous Agents

To understand exactly where the problem lies, we must distinguish between the three levels by which technology has advanced in recent years. The research divides AI tools into clear categories:

1. Autocomplete

This is the first wave, known for example from GitHub Copilot. These tools work on the principle of prediction – as you type, AI suggests the next line or an entire function. According to studies, this technology increases developer activity (measured by the number of "commits") by approximately 40%. It's a great helper for quickly writing routine code sections.

2. Sync Agents

The second generation, represented for example by Claude Code from Anthropic, works in real-time directly with the developer. They not only complete text but actively collaborate on file editing and problem-solving during the creation process. The cumulative effect of these tools increases productivity by 140%.

3. Async Agents

The newest and most powerful level is represented by autonomous agents, such as GitHub Coding Agent or systems built on OpenAI models. These agents receive a task (e.g., "fix a bug in authentication") and work on it independently, without the developer constantly having their hand on the keyboard. Their impact on the volume of code written is astonishing – they increase activity by up to 180%.

The "Bottleneck" Paradox: Why Doesn't Code Turn into Products?

Here comes the key finding that should interest every manager in a technology company, including those in the Czech Republic. Even though AI can generate a huge amount of code (commits), this growth drastically drops when we try to create a final product.

Research shows that while the volume of written code has massively increased thanks to AI, the number of new projects grows by only 50%, and the actual number of released versions of applications by only 30%. Why is this the case? The answer lies in the weak links hypothesis.

Software is not just text; it's a chain of steps: Writing $\rightarrow$ Code Review $\rightarrow$ Testing $\rightarrow$ Integration $\rightarrow$ Deployment. AI can extremely accelerate the first step (writing), but processes like review (code inspection by another human) and testing are still in human hands. If human capacity for review does not increase, additional code will only create a "pile of digital waste" that no one will have time to review and approve.

What Does This Mean for the Czech Market and Developers?

For Czech software companies and individuals, this result brings several fundamental implications:

Changing role of the developer: The role of "code writer" (coder) is changing to the role of "reviewer and architect." Your value will not be in how quickly you write a function, but in how effectively you can verify the code generated by an agent.
Quality vs. quantity: The market is starting to see an overload of low-quality applications. As App Store and Google Play analysis shows, the number of new applications is growing, but their overall engagement (usage) is stagnating. For Czech startups, this means that the path to success is not in "releasing more features," but in precise testing and solving real problems.
Affordability of tools: Most of these tools are globally available to Czech users as well.
- GitHub Copilot: approx. 10 USD/month (for individuals).
- Claude Pro (for Claude Code): approx. 20 USD/month.
- OpenAI API (for custom agents): pay-as-you-go.

In the context of the EU AI Act, it is also necessary to ensure that even if an agent writes code, legal responsibility for its safety and compliance with regulations remains with the human subject. Code automation must not mean automation of responsibility.

Conclusion

AI tools are incredibly effective at producing "material," but they still lack the ability to complete the entire productive cycle without human oversight. True productivity in 2026 will not be measured by the number of lines written, but by the speed with which we can safely and qualitatively integrate these lines into the real world.

Will AI replace developers because it writes code so quickly?

No. Research shows that AI creates a "bottleneck" in human oversight. Developers will spend less time writing syntax and more time on review, architectural design, and quality assurance. The role transforms, it is not eliminated.

Why then do we see so many new apps in the App Store that no one uses?

AI lowers the barrier to entry. It's extremely easy to create a basic version of an app (MVP), but AI cannot yet solve distribution, marketing, or deep user value that keeps people using an app. We have more "production," but not necessarily more "demand."

Is it safe to use autonomous agents for critical infrastructure?

Currently not without strict human oversight. Since AI can generate errors or security vulnerabilities (code hallucinations), a "human-in-the-loop" process is essential to meet safety standards and regulations like the EU AI Act.