As enterprises scale agentic coding tools, the study suggests the promised productivity is pooling at an unexpected stage of the pipeline — the humans who must vet every generated line.
AI coding agents are writing more code than ever. The trouble, according to a new study covered by Ars Technica, is that the volume of finished software isn't rising to match — the efficiency gains are getting stuck in human review.
The study's core finding is that productivity improvements from agentic coding tools are being 'absorbed' elsewhere in the development pipeline rather than converted into shipped software. Code generation was never the only constraint on how fast software gets built: every line an agent produces still has to be read, understood, verified and approved by a human engineer before it counts as delivered work. Once generation accelerates, the study finds, that review stage becomes the binding constraint — a 'bottleneck' that soaks up much of the time the agents were supposed to save. The result reframes a debate that has largely been fought over how fast agents can produce code, redirecting attention to the human step that follows.
The speedup is real. It just doesn't land where the ROI spreadsheets assume it does.
That distinction matters for anyone budgeting for agentic coding tools. If the constraint has shifted from writing code to reviewing it, then output volume — the metric vendors most readily showcase — overstates the actual gain in software delivered. A team can double the code its agents generate and still ship at the same pace, because approval capacity, not generation capacity, now sets the ceiling. Until review gets faster, the study implies, faster generation mostly means longer queues of unreviewed code.
Ars Technica's framing of the October 2026 study is blunt: the agents generate more code, but not more software.
萬安算交所 AIXX · 算力硬件实名会员制交易平台 · 进入货源大厅