Harvard researchers looked at practice in hundreds of companies and reached a hard conclusion: AI agents can produce code fast, but firms do not ship more software because of it. The gain at the writing stage is eaten by a bottleneck further down the pipe — human review.
What they counted
Fiona Chen and James Stratton used aggregated Jellyfish analytics: about 300 million work events — commits, pull requests, tickets — from more than 700,000 employees at 700-plus software companies from 2021 to March 2026. They timed the arrival of assistants (hints to a human) and agents (they write and open PRs themselves) from direct tool accounting and GitHub activity, then compared before and after across organisations.
After agents landed, average lines of code rise 30%, commits 20%, pull requests 23%. The share of closed tasks and epics in trackers such as Jira does not move statistically. Ticket difficulty does not shift either: companies did not start taking easier work to look more productive.
Where the pipe stalls
Time from opening a pull request to merge rises 49% on average. The share of requests sent back for rework almost doubles; comments per PR rise 35%. The share of staff doing review rises 14%. Headcount overall the researchers do not tie to AI: no mass layoffs and no hiring “for agents” in the data.
In theory AI could review itself. By March 2026 some form of automatic review existed at 80% of measured firms, but agents left only 23.3% of comments and touched 10.8% of pull requests. The lion’s share of checking is still done by people who do not trust raw agent code.
How to look at the numbers
The authors caveat: agents are still young, the data cuts off in March 2026, and 95% of firms in the sample already turned them on and are only learning where to let an agent in. The “write faster / review longer” ratio may shift. For now the picture is this: more code is not more software. Companies that score success by lines and PR counts are drawing themselves growth that is not there at the level of closed features.
