# AI Writes the Code Fast; Somebody Still Has to Read It > Source: > Published: 2026-10-09 19:43:50+00:00 There is a version of the AI-coding story that goes like this: the agents now write code so fast that software is nearly free, and the only remaining question is what engineers will do with all their spare time. [Kyle Orland at Ars Technica reports](https://arstechnica.com/ai/2026/10/ai-coding-agents-generate-more-code-but-not-more-software/) on a new study that suggests the answer is “code review, forever.” Harvard University researchers Fiona Chen and James Stratton drew on aggregated analytics from Jellyfish, a platform that measures engineering teams’ output: 300 million individual “work events” — commits, pull requests and the like — plus issue-management data covering more than 700,000 employees at over 700 software development firms, from 2021 through March 2026. That is about as close to a census of working programmers as this kind of research gets. To understand the finding, you need the plumbing. Code is not software. When an engineer — or an agent — finishes a change, it goes into a pull request, and somebody else has to read it, argue with it and approve it before it merges into the actual product. Imagine you run a restaurant and you buy a machine that lets your cooks plate dishes twice as fast. Nice. Except every plate still has to pass one health inspector at the pass, and the machine’s plates are, let’s say, inventive. Your kitchen’s output is now set entirely by the inspector. That is roughly what the data show: the average “review process” time between a pull request being submitted and being merged balloons 49 percent after AI agents are introduced. The granular numbers tell the same story from inside the review itself. After the AI agent shift, the researchers describe “the share of pull requests with changes requested nearly doubl[ing], and the number of comments per pull request increas[ing] by 35%.” So: more code gets written, and roughly commensurately more time gets spent sending code back with notes. More code, as the title of the research coverage neatly puts it, but not more software. Who does all this extra reviewing? People. The share of workers performing code reviews rose 14 percent after AI agents arrived. And for the “AI is coming for the jobs” ledger: cross-referencing total active workers in the Jellyfish data with LinkedIn records at those firms, the researchers write that they “cannot attribute significant employment changes to AI.” Nobody got replaced by the code-writing machine. A lot of somebodies got a new second job as its proofreader. The obvious rejoinder is that the machines should review their own homework, and the industry is trying: 80 percent of measured firms used some form of AI code review by March 2026. But AI agents were responsible for only 23.3 percent of all review comments and 10.8 percent of all pull requests. The tools that generate code at superhuman speed are, so far, doing a distinctly subhuman share of the checking. Humans remain responsible for the vast majority of both rejecting and accepting what the agents produce. The honest caveats matter. AI coding agents are young; their output has improved significantly even since the study’s March 2026 data cutoff. And while 95 percent of firms in the study have implemented agents by that point, plenty are still working out when an agent is the right tool and when it is a way to generate a very long pull request very quickly. The coding-time-versus-review-time trade-off could well improve as teams learn. Learning curves are real. So is the 49 percent. For now the trade looks like a genuine double-edged sword: gains in coding speed counteracted by increases in human review time and effort, at considerable expense in tooling to boot. There is an old lesson in here about bottlenecks — automating the step that wasn’t the constraint mostly produces a taller pile in front of the step that is. The machines have learned to write. The reading, it turns out, is still on you.