October 9, 2026
AI agents write more code, but releases rise at most 30%: human review eats the gain
Autonomous agents raised commits by 240%, but projects and releases only by 80% and 30%. The NBER paper's authors see the bottleneck in human review: AI does not replace it, it runs into it.
The sharpest jump comes from autonomous agents: they carry a task through on their own, without constant supervision. Commits from them rise 240%, projects up to 80%, releases up to 30%. The closer to release, the less of that growth is left.
The paper compares three waves of tools. This is a narrow measurement that the authors raise to a rule: the gain in commits gets stuck at the stage where a human checks and reworks the code.
Three waves. Autocomplete gave +30% commits. Interactive agents, which work with a human in dialogue, gave +180%. Autonomous ones gave +240%.
Weak link. The elasticity of substitution between AI output and human review and rework is 0.23. The authors treat AI and humans as complements: growth in AI output does not replace the human stage, it runs into it. This is their ‘weak link’ hypothesis.
What it means for builders. For those building with agents, the bottleneck has shifted: writing code got faster, but accepting it still falls to a human.
The number of apps on four major app marketplaces grew, but downloads in the same cohorts did not increase.
The NBER paper came out May 28, 2026, DOI 10.3386/w35275.