Autonomous AI triples GitHub commits, but real releases rise just 30%

Generating code has officially outpaced shipping it. A working paper from the National Bureau of Economic Research (NBER) by economists Mert Demirer, Leon Musolff, and Liyuan Yang analyzed telemetry across more than 500,000 developers on GitHub. Their findings demonstrate where the productivity promise of artificial intelligence actually vanishes.

Across tool generations, the surge in raw code output is massive:

• Autocomplete lifted commits by roughly 30%.
• Interactive coding assistants boosted them by 180%.
• Fully autonomous agents drove commits up by 240%.

Downstream, however, the compounding gains evaporate. For autonomous agents, that 240% surge in commits translates to an 80% increase in active projects, which collapses to just a 30% gain in actual shipped releases.

The human bottleneck

The authors explain this drop-off through weak-link theory: overall throughput is bounded by the slowest stage in the chain, which remains human evaluation. The study calculated an elasticity of substitution between AI and human labor of just 0.23, reflecting heavy complementarity rather than direct replacement. More automatically generated code simply forces more manual code review, context matching, and system integration.

A flood of finished artifacts does not guarantee consumption, either. Examining the Apple App Store, Google Play, Chrome Web Store, and SourceForge, the researchers noted a sharp increase in published apps, yet aggregate consumer usage remained virtually flat.

The real cost of autonomous delivery

A benchmark test published in April 2026 by cruxevals illustrated why autonomous development stalls outside the text editor. An AI agent running OpenClaw on Claude Opus 4.6 was tasked with building and releasing an iOS app, Breathe Easy. While writing and packaging the application took 45 minutes and $25 in compute, platform overhead told a different story:

• Environment setup consumed eight hours over two dry runs.
• App Store review queues took nine days to clear.
• Autonomous status polling during the waiting period burned $975 in tokens.

The agent also fabricated a contact phone number that slipped past review, dropped its credentials, generated store screenshots with visible defects, and shipped a non-functional audio toggle. Writing syntax may take minutes, but verifying edge cases, navigating platform bureaucracy, and finding real users remains the unyielding constraint.

Source: habr

Study: NBER Working Paper 35275

Case: cruxevals