Since we've set up early projects built end-to-end by an agentic 'harness'—and a dedicated 'agentic digital factory' within our Michelin teams—I've seen merge request volume truly exploding. Codex alone now writes up to 30% of our total code—even though only a handful of engineers, out of roughly 200, actively drive it. Huge volume of MRs. A lot of noise. High 'RPM', as we would have said of an engine.
A new NBER paper just put numbers on that same feeling. Researchers tracked more than 100,000 GitHub developers across three generations of AI coding tools—autocomplete, synchronous agents, autonomous agents—and found cumulative commits up 40%, 140%, and 180%.
Impressive?
Except that 180% falls to 50% once you count actual projects shipped, and to 30% once you count releases. Across major app marketplaces, the same pattern: more apps created, usage barely moving.
The NBER paper calls this the "weak-link hypothesis": gains at one stage of a production chain get absorbed by the human-bottlenecked stage next in line. It's Amdahl's Law for organizations: your overall speedup is capped by the part you didn't accelerate. Review. Architecture. Security. Product decisions. Adoption. None of that got faster just because commits or MRs did.
This is exactly what I was starting to feel a few weeks ago in this short blog piece The Engine Was Ready. The Infrastructure Wasn't., exploring the analogy with thermic-to-electric automotive transition.
An electric motor doesn't just replace a combustion engine—it changes the physics of the whole vehicle. A combustion engine only makes power in a narrow RPM band, which is exactly why it needs a gearbox in the first place. An electric motor makes full torque from zero, so the gearbox—the clutch, half the transmission—simply stops being necessary. Coding agents do something similar to the "gearbox" of software delivery: the tickets, the sprint ceremonies, the sequential handoffs that mostly existed to translate work into pieces a human could carry. Extraordinary acceleration, almost no friction producing the first version of anything.
But acceleration isn't transportation. Roads still matter. Charging infrastructure still matters. A faster engine doesn't raise the average speed of a journey if everything downstream of it hasn't been rebuilt to match.
Most organizations measuring AI's impact right now are just reading "engine RPM"—lines of code, MRs, velocity points. But the business only cares about kilometers traveled, doesn't it?
This being said, we shouldn't dismiss the magnitude of what's happening across the development ecosystem. The right posture is to play the long game, the way the electric transition played out: progressive hybridization first, and, more than anything, sustained investment in the infrastructure around it—harness, upskilling, model routing. The power grid, in short. That alone turns acceleration into a real step change in productivity and ROI.
