More code, not more software: review is the binding step.更多代码，不等于更多软件：瓶颈在人工评审这一步。

AI coding agents now add roughly 30% more code at the firms that adopted them — yet no measurable increase in shipped software, because the human reading the code, not the writing of it, is what caps delivery.

A Harvard working paper by Fiona Chen and James Stratton (August 2026) tracked about 300 million work events at 718 software firms and some 726,000 workers from 2021 to March 2026, using Jellyfish engineering data. After teams adopted coding agents in the Claude Code, Cursor and Devin class, each engineer added about 1,494 extra lines a month on a 5,017 baseline, with 20% more commits and 23% more pull requests.

Finished work did not follow. Jira issues and epics resolved barely moved: plus 0.12 on a 3.67 baseline, statistically indistinguishable from zero.

Here is where the time went. Pull-request open-to-merge rose 49%, about 3.45 days on top of a 7.03-day baseline. Formal change requests nearly doubled, up 12 points from 13% of PRs. Comments per pull request rose 35%. Firms even pulled 14% more people into reviewing.

The mechanism is simple once you see it. Generating code got cheap, so agents emit more of it. Reviewing code did not get cheap: one person can carefully judge far fewer pull requests than an agent can open. After AI-written commits caused outages, some firms added senior-approval rules — more review, not less. Review bots did not absorb the load either; only 10.8% of pull requests got even one AI comment.

The second-order effect: when new code is nearly free, the incentive is to open more and bigger PRs, while reviewer attention stays fixed. The faster tool can make shipping slower. The agents made writing code cheap; shipping software still costs a human reading it.

The strongest counter: this is a job-market paper, not peer-reviewed, and firm-level correlation is not a clean experiment. Task-level speedups are real — LinearB's data on 8.1 million PRs shows AI-assisted pull requests clearing review faster once started, 3.2 hours against 4.2. That wins where review capacity is elastic. Our claim is narrower: firm-level output over quarters, where the human gate is fixed — and LinearB's own 30-day merge rate, 32.7% against 84.4%, points the same way.

If you lean on coding agents, stop grading them by lines or PR count. Track cost and time per resolved issue that actually ships, and cap PR size so review doesn't quietly become your bottleneck.

This breaks if an independent or peer-reviewed study shows firm-level shipped output rising with agent adoption, or merge times returning to baseline, by June 30, 2027.