More Code, Same Software
Henry Kim ·
Harvard researchers analyzed engineering records from 718 companies, roughly 300 million work events. After adopting AI coding agents, lines of code rose 30%, commits 20%, pull requests 23%. But completed software did not rise with them. The resolution rate of Jira issues and epics showed no statistically significant change. The study is by Harvard's Fiona Chen and James Stratton, using data from engineering analytics firm Jellyfish, covering GitHub activity and Jira records from 2021 through March 2026. The paper is titled 'Artificial Intelligence in the Firm: Bottlenecks in Software Production.' It is observational, so causation is not proven. But the scale makes it hard to dismiss. Something else rose instead. Average pull request review time went up 49%. The share of PRs sent back for changes nearly doubled. The share of workers doing code review rose 14%. The bottleneck moved from writing to review. This is not a coding story. It is the same story as last week's. Not which tool you picked, but what happened after adoption. Agents made code cheap. But a software company does not sell code. It sells changes that are reviewed, tested, and integrated. The machine saved implementation time and created verification labor downstream. The obvious response is AI reviewing AI, and companies are doing it. By March 2026, about 80% of measured firms used some form of AI code review. The review wall still stands. Agents produced 23% of review comments and touched 11% of pull requests. The rest is still human. GitHub even built ReviewBench, a benchmark for whether AI reviewers catch meaningful bugs or flood pull requests with low-value comments. The bigger generation gets, the bigger this problem gets. Why? A good reviewer knows architecture, not just syntax. Customer expectations, operational history, undocumented dependencies. The reviewer knows which strange-looking behavior is actually intentional. An AI reviewer catches obvious bugs and misses the one architectural mistake that matters. Or it does the opposite: a wall of plausible, low-priority comments that makes the human slower. GitHub engineers shared a telling episode. Giving Copilot code review better general-purpose tools made it worse. Review cost went up, issues found went down. Rewriting the agent's instructions around the actual review workflow cut review cost by roughly 20%. Evidence that workflow design, not model capability, decides productivity. The prescription is not to generate less code. It is to make each unit of generated work easier to verify. Keep agent tasks narrow. Write acceptance criteria so machines can read them. Let tests and static analysis reject obvious failures before a human sees the PR. Draw lines agents may not cross around architectural decisions. Measurement has to change too. When code is nearly free to produce, lines and commit counts lie more each month. Measure something else. Issue cycle time, deployment frequency, rollback rate, reviewer hours, production-qualified output per engineer. The conclusion is narrow and solid. AI made code generation cheap, which made review, integration, and organizational judgment the scarce resources. The winners will not be the teams whose agents write the most code. They will be the teams that establish trust in each change with the least human attention. What is your team measuring right now. Lines of code, or shipped software.