The Bottleneck Moved Floors

The bottleneck for shipping production-grade software just moved floors. Two years of debate centered on model capability. The work has migrated upward into the layer that decides what code is supposed to do, who reviews it, and how the intent gets captured before any agent writes a line.

Dex's essay in HumanLayer's repository gathers the signal. Matt Pocock's line that "codebases are falling apart faster than they ever have before" reads as alarm, but it points at a specific mechanism. The Faros AI industry report finds that since January and February, when AI coding tools went mainstream, pull-request review volume is up, more PRs merge without any review, incidents are up, and bugs per developer are up. The model is not the choke point. The layer above it is.

The misreading is the popular one. "AI is breaking code" implies the model is regressing. The mechanism is what engineers already know: when a system produces more output per unit of time, the constraint migrates to whatever checks that output. StrongDM's lights-off factory, OpenAI's Symphony talk, and Mario's slowdown call at AI Engineer Europe describe the same migration from different angles.

One inference: at companies scaling AI coding, the next dollar, by most accounts, earns more by going into review, specification, and intent-capture than into another model swap. Confidence is correlation-strong, not causal-settled.