AI coding agents boost code output but stretch review times 49 percent, study finds

More code, slower merges

AI coding agents are reshaping engineering work rather than simply speeding it up, according to Ars Technica's report on new research. The study, which draws on data from Jellyfish, found that the average time between a pull request being submitted and merged grew by 49 percent after teams introduced AI agents. The report's central conclusion is that agents generate more code without producing more software, and the review process explains the gap.

What the numbers show

The research, credited by Ars Technica to Chen and Stratton, is based on data running through March 2026. By that point, 95 percent of the firms studied had adopted AI coding agents, so the findings describe mainstream usage rather than early experiments.

The granular figures show where the friction sits. The share of pull requests in which reviewers requested changes nearly doubled after agents arrived, and comments per pull request rose by 35 percent. Agent-written changes, in other words, tend to arrive quickly but need more rounds of correction and discussion before anyone is willing to merge them.

Reviewers absorb the extra load

Teams responded by shifting people toward review work. The researchers recorded a 14 percent increase in the share of workers performing code reviews once agents were in place. At the same time, they write that they cannot attribute significant employment changes to AI, based on total active worker counts across Jellyfish and cross-referenced against LinkedIn data at the same firms. The overall picture is a redistribution of effort — less time writing code, more time judging it — with headcounts roughly stable.

AI review tools have not caught up

Automation could in principle absorb the review bottleneck, but the data says it has not yet. By March 2026, 80 percent of the measured firms used some form of AI code review, yet AI agents accounted for only 23.3 percent of review comments and were responsible for just 10.8 percent of pull requests. Humans still handle the bulk of the evaluation that agent-generated changes now demand, and the researchers describe AI's contribution to reviewing so far as marginal.

A fast-moving target

Ars Technica adds a caveat: the study's data ends in March 2026, and agent capabilities have kept changing since. Many firms are presumably still learning which tasks suit agents and which do not, and the researchers suggest the trade-off between coding time and review time could improve as teams build that experience.

Why it matters

For engineering leaders, the study is a warning against reading raw output metrics as productivity. Pull request volume and code generation can climb sharply while the number that matters — how quickly verified change reaches the codebase — moves the other way. Teams adopting agents should budget reviewer capacity as deliberately as licence costs, and track merge latency and change-request rates alongside generation volume.

The findings also point at where tooling goes next. If writing code is becoming cheap and reviewing it is becoming expensive, the bottleneck for software delivery is shifting to evaluation — and that is where engineering processes and products are likely to be redesigned next.