AI agents write more code, but there is little increase in finished software

After deploying AI agents, companies produce 30% more lines of code and close tasks at the same pace as before. A Harvard study shows that the gains are eaten up by code review. This points to where managers should measure AI's impact.

AI coding tools are usually sold with one promise: the team will do more in the same amount of time. Data from more than 700 software companies shows that this promise is only half fulfilled. There is indeed more code, but finished software increases much more slowly, or not at all. For a manager paying for licenses or for the hours of a software house, that is an important difference.

What the data from more than 700 companies showed

Fiona Chen and James Stratton, economics PhD students at Harvard, described this in the paper “Artificial Intelligence in the Firm: Bottlenecks in Software Production”. They used data from Jellyfish, a platform that measures the work of engineering teams. The sample includes 300 million work events, including commits (saved changes in code) and pull requests (change submissions to be merged into the main codebase). They also used data from task management systems such as Jira. The study covered more than 700,000 employees at 718 companies from 2021 to 2026, through March of this year.

The researchers compared companies before and after deploying AI tools. They also distinguished between two types of tools. Assistants suggest code that is still mainly written by humans. Agents independently write and submit code based on prompts.

After agents were introduced, the number of lines of code rose by an average of 30%, commits by 20%, and pull requests by 23%. The pace of closing tasks and entire features tracked in Jira did not change in a statistically significant way. Tasks also did not become larger or more complex, so this difference cannot be explained by more difficult work.

The bottleneck is the human reading the code

The gains from faster writing disappear at the code review stage, when another programmer checks changes before they are merged. The time from submitting a pull request to its merge increased by an average of 49%. The share of submissions where the reviewer requested fixes nearly doubled. There are 35% more comments per submission. Companies responded in their own way: the share of employees focused on code review increased by 14%.

AI could theoretically also help with review, but for now it does so only to a limited extent. By March 2026, as many as 80% of the surveyed companies were using some form of AI in code review. However, agents accounted for only 23.3% of review comments, so most of this work is still done by humans. The researchers also found no significant changes in employment. They checked this in Jellyfish data and LinkedIn data.

These results should be treated cautiously. Agents had already been deployed by 95% of the companies in the sample, and many of them are still learning what to use them for. The tools have changed significantly since March. So the balance may improve, but for now there is no data showing that.

Other studies show the same pattern

Other researchers reached similar conclusions. Leon Musolff of Wharton and Mert Demirer and Liyuan Yang of MIT analyzed the work of more than 100,000 programmers on GitHub from 2022 to 2026. According to them, autonomously operating agents increase coding productivity by a total of 180%. However, the number of released software versions increased by only 30%, and completed projects by 50%. Musolff notes that 30% more releases is still a result that few technologies deliver. At the same time, the authors write that the main constraint shifts from writing code to reviewing, integrating, and distributing it. In the Apple App Store, the number of new apps increased from around 30,000 per month before 2025 to around 100,000 in April 2026, but app usage is flat or declining.

DX Research, a company studying developers' work, observed 4.2 million developers from November 2025 to February 2026. About 93% of them use an AI assistant at least once a month. However, the number of completed pull requests increased by only about 10%. The reported time savings plateaued at about four hours per week. Laura Tacho, Chief Technology Officer (CTO) at DX, says plainly that this is a management problem:

The most uncomfortable result came from an experiment by METR, an organization that studies AI safety. In a study published in July 2025, sixteen experienced open-source programmers worked 19% more slowly with AI than without it, even though they themselves believed they had sped up by 20%. This is a small sample and a different mechanism from the Harvard study. Still, it shows that team perceptions can be a poor metric.

Where to look for gains from AI instead of counting lines of code

A simple conclusion follows from this data: the number of generated lines says little about value for the company. If the gains are eaten up by review and verification, that is where attention and money should be invested. Startups are already addressing this gap. Zurich-based Klarent announced on October 9 an early-stage seed funding round of €7.13 million to develop autonomous testing. Its agents write tests, but the tests themselves run in a repeatable way under human supervision. The client pays only for verified results. We know the customer list, which the company says includes JD Sports, NZZ, and Sixt, only from its own announcement. There is also no independent data showing that such tools shorten code review.

A second direction is pointed out by Itay Sagie, a strategic advisor writing for Crunchbase News. In his view, the advantage is shifting to places where AI performs repetitive work inside the customer's processes. As an example, he cites, among others, Schneider Electric's announced October 5 acquisition of PTC, a provider of software for designing, manufacturing, and servicing products, for about $22.6 billion. This is an opinion, not a research finding. In our view, however, it fits the Harvard data: value is visible where AI completes the task, not where it adds work for the next link in the chain.

How to talk about AI with a software house

If a contractor claims that AI will help deliver more, it is worth asking about end metrics rather than writing speed. These questions can help:

• how many features reach users each month before and after deploying agents,
• how long code review takes and how often submissions are sent back for fixes,
• who tests code written by AI and how,
• how the hourly rate changes if part of the work is done by a machine.

A good partner will answer these questions with data. If, in response, you hear about the number of lines of code or developers' feelings, studies from Harvard, Wharton, and DX show why that is not enough. AI in programming works, but for now it shifts work from writing to checking. Companies that rebuild review and testing have a chance to see real gains.