AI Coding Agents Write 30% More Code – and Resolve Zero Additional Issues, Harvard Study Finds

TEXT START: Enterprise technology leaders are currently caught in a measurement trap.

The Dissection

The text performs two operations. First, it correctly separates activity from completed economic output: more code, commits, and pull requests produce longer review queues, more revisions, and no statistically significant increase in resolved issues. AI accelerates the upstream task while exposing review as the constraint.

Second, it domesticates the threat. The problem is framed as a workflow defect solvable through better testing, review tools, and organizational redesign. AI adoption itself is treated as inevitable and desirable; only the plumbing supposedly needs repair. The article identifies a jammed transmission, then assumes the human transmission can be upgraded indefinitely.

The Core Fallacy

The central error is treating bottleneck migration as a temporary management problem rather than a selection mechanism.

Under the Discontinuity Thesis, AI making code generation cheaper does not preserve software labor. It makes the remaining human bottleneck more visible, expensive, and vulnerable. Review is not proof that humans remain structurally necessary. It is the next queue to automate, narrow, standardize, or eliminate.

The statement that human oversight remains an “absolute requirement” converts a present condition into a permanent law. Even the text names automated testing and improved review tooling as ways to reduce that requirement. The current productivity plateau therefore shows that firms have not yet rebuilt the pipeline around machine production—not that engineers have escaped obsolescence.

Hidden Assumptions

• More review capacity can be purchased without changing employment levels or bargaining power.
• Human code review is a durable moat rather than a temporary lag defense.
• Firms will respond by adding tooling and staff instead of restricting output, redesigning jobs, or cutting headcount.
• Jira issues and completed epics fully capture software value.
• Flat issue resolution means AI has failed, rather than indicating that the old production system cannot absorb cheap machine output.
• The generator-reviewer division will remain stable instead of review itself becoming increasingly automatable.
• Productivity gains must appear as more completed work rather than fewer workers needed for the same work.
• Competitive firms will pay humans to manage an expanding “babysitting tax” indefinitely.

Social Function

Classification: partial truth, transition management, and ideological anesthetic.

It is not pure copium. The bottleneck, review burden, and gap between code volume and business output are real within the text’s evidence. But the article converts a displacement signal into a process-improvement agenda. Its message to management is: continue adopting AI, purchase complementary capacity, and redesign workflows. It avoids the harder implication that once downstream controls are automated, the same bottleneck that temporarily protects senior engineers becomes the next target of substitution.

The Verdict

This is a useful autopsy of first-wave AI deployment and a weak forecast of labor’s future. It shows P1 colliding with obsolete workflows: AI makes code cheap, while humans become the queue. That queue is lag, not salvation.

The article mistakes a clogged transition for a permanent limit. Firms will either automate review, constrain machine output, or use the new throughput potential to reduce human staffing. The immediate absence of additional resolved issues does not rescue software workers. It proves that the enterprise has not yet finished rewiring itself for machine production.