By following more than 100,000 GitHub developers, three NBER economists show that the spectacular gains of code agents evaporate as we move up to the delivery stage.
In a NBER working paper published in May 2026, Mert Demirer, Leon Musolff and Liyuan Yang cross-referenced the activity of more than 100,000 GitHub developers with telemetry data from the use of AI tools, according to a protocol for the study of paired events. Their objective: to distinguish what AI saves at the level of the task from what it saves at the level of the delivered product.
At the code level, the progression faithfully follows the generations of tools: about 40% more commits with autocomplete, 140% with interactive coding agents, 180% with autonomous agents. But the effect erodes as you go up the production chain: the 180% observed on coding activity falls to 50% on the number of projects, then to 30% on actual software releases.
The authors attribute this attenuation to human links in the chain, with an estimated elasticity of substitution of 0.25 between human effort and AI effort — a sign of strong complementarity rather than replacement. A check on four major application marketplaces confirms the diagnosis: a few more new applications, but no increase in overall usage.
Read the full article on the original source
Go to the source →