VQV Signal

RESEARCH WATCH TECHNICAL

Harvard working paper on 720,000 engineers links AI coding agents to slower code review

A post on X by @konstantinsaifo highlights a Harvard working paper by Fiona Chen and James Stratton that uses Jellyfish data on about 720,000 workers at 655 companies. According to the post, with AI coding agents, lines of code rose 44% and pull requests rose about 36%, but finished software rose only about 14%. The time from review submission to approval grew 56%. The author frames the pattern through Goldratt's theory of constraints, in which speeding up one step does not by itself speed up the whole system.

Source: Scobleizer News Pulse (X) · x.com Found via Robert Scoble’s Scobleizer News Pulse · a post on X, not independently verified Published 2026-11-19T20:40:00+00:00 Detected 2026-11-20T05:01:43+00:00
View original source

A post on X by @konstantinsaifo highlights a Harvard working paper by Fiona Chen and James Stratton that uses Jellyfish data on about 720,000 workers at 655 companies. According to the post, with AI coding agents, lines of code rose 44% and pull requests rose about 36%, but finished software rose only about 14%. The time from review submission to approval grew 56%. The author frames the pattern through Goldratt's theory of constraints, in which speeding up one step does not by itself speed up the whole system.

AI-assisted summary based on the listed source.

Konstantin Saifoulline highlighted a Harvard working paper by Fiona Chen and James Stratton that uses Jellyfish data on about 720,000 workers at 655 companies. With AI coding agents, lines of code rose 44% and pull requests about 36%, but finished software rose only about 14%, while the time from review submission...

For engineering teams adopting AI coding agents, the findings suggest code review may limit output, not code generation alone. Engineering managers and tool buyers may need to track review capacity alongside code volume. The source is a social post about a working paper, so the full methodology and findings should be checked before drawing firm conclusions.

Signal Strength 91% Technical label WATCH Public Interest 26 Category RESEARCH Reader Depth TECHNICAL

Signal Strength reflects source quality, relevance, freshness and evidence. Public Interest helps organize discovery; it is not proof of truth.

Public Interest components
Recognizable Entity Score 0 Practical Impact Score 8 Novelty Interest Score 70 Consequence Score 30 Curiosity Score 16 Shareability Score 42

VQV surfaced this signal because it is recent, relevant to AI Coding Tools, connected to Scobleizer News Pulse (X).