Why AI Coding Agents Can Write More Code but Deliver Less Software

Why can AI write more code without delivering more software?

Writing code is only one part of making software. The code has to fit with the rest of a project, be checked for mistakes, revised when needed, and incorporated into a working product.

AI coding agents are programs that can pursue tasks using software tools and take actions with some autonomy. ⁵ They can help produce code, but that does not automatically remove the work needed to assess and integrate it.

A study based on engineering activity across hundreds of firms found that coding gains were absorbed by later steps in the production process. The researchers reported longer code reviews, more reviewer comments and a greater likelihood that proposed changes needed revisions. ¹

What did the research actually measure?

Harvard researchers Fiona Chen and James Stratton used aggregated analytics from Jellyfish, a company that measures engineering-team activity. The data included 300 million work events, such as commits and pull requests, along with issue-management data. It covered more than 700,000 employees at over 700 relevant firms. ¹

A pull request is a proposal to add or change code in a shared project. It gives colleagues a chance to review the proposed change before it is incorporated.

The reported finding is not that AI never helps an individual programmer. It is that, in the data described, faster code production did not clearly translate into more overall software output or lower employment. ¹ The provided source summary does not give enough detail to assess every method or limitation of the study.

What makes review a bottleneck?

When more code arrives for review, people still need to decide whether it is correct and suitable for the project. If proposed changes need revisions or draw more comments, reviewers and developers spend time resolving those issues. ¹

That work can limit the benefit of faster code generation. The study describes a production process in which efficiency gains at one step were taken up by constraints later on. ¹

This does not mean every AI-generated change is poor or that review is wasted effort. It means that code generation speed by itself is not a complete measure of whether a team is delivering more finished software.

Are AI coding tools reducing software jobs?

The available reports do not support a simple conclusion that software developers are disappearing. CNBC reported that U.S. software-development job postings had risen since February 2025, while overall job postings fell. It also described the rebound as led by senior and AI-focused roles, alongside weaker entry-level hiring. ²

Job postings are not the same as hires, employment totals or long-term career prospects. The report suggests that demand may be changing, but it does not establish that AI caused the changes or show what will happen next.

The research on coding output also found little evidence that firms increased software output or reduced employment through the use of AI coding tools. ¹ That is a finding from the data and period studied, not a guarantee about every company or future hiring.

What should readers take away?

For people using these tools, the practical point is to distinguish code produced from software successfully delivered. Review and revision remain part of the work described in the study. ¹

For people watching the job market, both developments matter: AI agents are appearing in more coding work, while software-development postings have recently risen. ² Neither fact alone settles whether developers will be needed in greater or smaller numbers over time.

The evidence supports a measured conclusion: AI can speed up code generation, but the time saved may be offset by review and rework. The sources provided do not establish how large that effect is across all teams or how it will change.