The explanation for that discrepancy may be discovered immediately within the code evaluation course of, which takes markedly longer on common after the introduction of AI coding brokers. General, the typical “evaluation course of” time between a pull request getting submitted and it being merged into the codebase balloons 49 % on common after AI brokers are launched. That impact may be seen in additional granular information, too, with “the share of pull requests with adjustments requested almost doubl[ing], and the variety of feedback per pull request increas[ing] by 35%” following the AI agent shift, the researchers write.
In response to this modification, the researchers discovered a 14 % enhance within the share of employees performing code evaluations after AI brokers’ introduction. Additionally they write that they “can’t attribute important employment adjustments to AI” after complete energetic employees throughout Jellyfish and cross-referencing with LinkedIn information at these companies.
Credit score:
Chen and Stratton
Whereas AI might additionally theoretically assist with this evaluation course of, the researchers discovered that, to date, that affect has been marginal. Though 80 % of measured companies used some type of AI code evaluation by March 2026, AI brokers have been solely liable for 23.3 % of all evaluation feedback and 10.8 % of all pull requests, suggesting people have been nonetheless liable for the overwhelming majority of this work.
AI brokers are nonetheless a comparatively new a part of the coding world, in fact, and there have been important updates and upgrades to their output even since this examine’s March 2026 information cutoff. And whereas 95 % of companies within the examine have carried out AI coding brokers by this level, many are doubtlessly nonetheless going by means of a studying course of relating to when and methods to greatest deploy them. These sorts of “coding time versus evaluation time” trade-offs might enhance as software program engineering groups acquire extra expertise with the professionals and cons of siccing an AI agent on specific coding issues.
For now, although, letting AI write your code seems like a double-edged sword, with will increase in coding velocity counteracted by comparable will increase in human code evaluation effort and time. It’s the type of consequence that makes us marvel if the appreciable time and expense to get AI coding brokers working is admittedly price it for many firms.
