Type: url
Original: https://arxiv.org/pdf/2607.01087
Status: generated
Summary
A first-person case study in which a single expert engineer used frontier AI coding agents over 12 weeks to build a document-accessibility remediation system, leaving an empirical record of 88 contemporaneous field notes, 18,662 commits and roughly 1.6 million lines of active artifacts. From that record the authors induce a candidate middle-range theory, “cheap code, costly judgment”, whose central process is governance conversion: agentic velocity exposes structural failure classes, human architectural-contextual judgment interprets them, and those failures are converted into durable, machine-actionable governance that narrows the action space of subsequent agent work. The paper argues this ex-post, failure-induced discovery of controls complements the ex-ante governance models in prior work, and that under agentic velocity the scarce human work shifts from implementation-level review to problem framing, abstraction discovery and architectural judgment.