Huntley’s pre-commit gate versus instructions that drift

October 13, 2026 primary account: Geoffrey Huntley, “nobody should type code by hand anymore. here's what's left.” A four-part account of a conversation with Josh Wulf, not a controlled evaluation. He argues that the scarce human work is vision, explaining the system and defining verifiable properties, rather than typing and eyeballing each line. In his own diagnostic-library refactor, a later agent loop stopped using the desired module; rather than updating AGENTS.md, he says he added a pre-commit diff hook that rejects direct calls to observability modules and returns actionable feedback. Willison excerpted this particular passage. It is an appealing example of making a rule executable at the moment an agent tries to commit, but the reported ‘ten out of ten’ is not accompanied by a sample, tests or publication of the check’s coverage. A hook can catch a known direct-call pattern; it cannot infer whether every path uses diagnostics correctly or whether the diagnostic design serves the requester.

Provocation to retain, not an established general rule. Huntley says he pushes to main without PRs in his greenfield/personal work and runs agents with sudo on a rebuildable NixOS machine, under continuous system-level tests. He contrasts that workflow with domains moving money or other higher-stakes systems, where he says caution is needed. The claim that conventional PR review and manual code are obsolete is his stance, not supported by paired reviewer effort or customer/repair outcomes in this essay. Compare Anthropic, which retains human intent review despite high change throughput, and Jain’s behavioral harness, which does compare a held-out task pass but not deployment. Useful future counterpoint when someone equates review with reading each line: ask who authors the check, who sees a missed behavior, and how a change is blocked before deploy.