A new study tracking firms on the Jellyfish engineering platform finds that AI coding agents have achieved exactly what they promised and exactly what skeptics feared — simultaneously. Code gets written faster. But the code that gets written requires dramatically more human oversight to land, and the net effect on actual shipped software is a wash. The headline number: a 49 percent increase in average review-process time after AI agents are introduced. Pull requests submitted by agents need changes requested at nearly double the previous rate, and review comments per PR climb 35 percent. The AI writes fluently and prolifically; it just doesn't write well enough to trust without extensive human checking. The workforce response is revealing. Firms shifted 14 percent more workers into code review roles after adopting AI agents. The researchers cross-referenced Jellyfish data with LinkedIn and found no significant employment changes attributable to AI — nobody got fired, but labor got redirected from creating to inspecting. That's a reallocation, not a reduction. AI-assisted code review exists but hasn't closed the gap. By March 2026, 80 percent of measured firms used some form of AI review tooling, yet AI agents produced only 23.3 percent of review comments and handled just 10.8 percent of pull requests. Humans still do the vast majority of quality gatekeeping. The machine generates; the human validates. That ratio hasn't budged meaningfully. The structural problem is familiar from manufacturing: automating the easy part of a pipeline doesn't speed up the pipeline if the hard part stays manual. Code review is cognitively expensive — it requires understanding intent, architecture, edge cases, and security implications. Those are precisely the capabilities AI agents lack at production quality. Speeding up code generation without proportionally speeding up review creates a pileup, not a speedup. The study's March 2026 cutoff means newer agent capabilities aren't captured, and the researchers acknowledge a learning-curve effect. Teams are still figuring out when to deploy agents and when not to. But the pattern is structurally important: tools that generate volume without quality shift costs downstream rather than eliminating them. That's extraction dressed as efficiency — vendors capture subscription revenue while engineering teams absorb the review burden. For now, AI coding agents look less like a productivity revolution and more like a workflow rearrangement. The code gets written somewhere else; the work of making it shippable stays with humans. Companies paying for agent licenses should be measuring cycle time from commit to merge, not just lines generated.
