Real-world scaling cases of coding agents: Lovable, Codex, Claude Code, SWE-1.7. Token costs, PR governance, risk classification, and what it means for engineering productivity.
Lovable reported $85K in tokens since January to scale from 20-30 to 150+ pull requests per week. Simon Willison spent $149.25 in tokens to write sqlite-utils 4.0 primarily with AI. Cost varies depending on volume, model, and task complexity.
Lovable classified changes by risk level, implemented layered automated review, and created reusable skills. The principle: define where humans step in and where the machine decides on its own, with impact classification for each change before approval.
Data from the Ramp Economics Lab shows that companies with intensive AI adoption hire more, not fewer. Coding agents amplify productivity, but human attention remains the scarcest resource in the process. The gain is in freeing people for higher-value decisions.