Harvard study of 718 firms finds AI coding boosts individual output ~30% but firm-level shipping doesn't move
DavidLinthicum · x · 2026-10-08
A rigorous study by Harvard economists Fiona Chen and James Stratton, using 300 million work events across 718 firms, finds AI coding agents boost individual coding output 30% (assistants 10%) — yet firm-level software output and employment don't budge. The gains don't ship.
Why: code review becomes the bottleneck. After agent adoption, review times jumped 49% and the share of PRs requiring changes nearly doubled. The tools shift the constraint from writing code to reviewing it; without process changes, individual productivity gains never reach shipped software.
More from coding & agent
- Gateway logs miss direct agent calls: the real gap in AI agent audit trails — Exotic-Border-5328 · 2026-10-08
- Awesome-ai-apps: 132 Free Open-Source AI Agent and RAG Projects, 16.1k Stars on GitHub — Arindam_1729 · 2026-10-08
- H-Company optimizes VLM serving for computer use agents with NVIDIA Dynamo — NVIDIAAI · 2026-10-08
- Same game, same prompts, 3 engines: hands-on comparison building with Claude Opus — rubenrb02 · 2026-10-08
- Qwen 3.8 Flash Next feels like 'Claude 4.6 at home' for three.js coding, Reddit user reports — hiImMate · 2026-10-08
- go-harmony: a developer's algorithmic take on music theory, plus a forthcoming book — jedisct1 · 2026-10-08