High-stakes procedures are often the accumulated record of past failures: ambiguous language gives way to rigid checklists, controlled vocabularies, and verifiable evidence. The Navy’s fuel-line rules followed the Lexington’s catastrophic explosions; aviation checklists followed a B-17 crash, and strict runway phraseology followed the Tenerife collision. HACCP, audit controls, and double-entry bookkeeping similarly turn hard-won judgment into repeatable rules. The central claim is that agents are accelerating this “crystallization”: as automated work expands, organizations must convert tacit decisions into machine-checkable constraints, evidence trails, and explicit points where human judgment is required.
Examples from recent engineering and finance show both the gains and the gaps. A coding team replaced diff-centered reviews with typed sections for intent, criteria, invariants, decisions, and evidence, cutting review time from five days to under three hours. Another derived review rules from 33 production incidents. But a poll of roughly 60 engineers found that, in one team’s 95 architecture decisions, only about 5 percent were enforced by lint rules, 55 percent by tests, and 40 percent remained prose; roughly one-fifth were effectively unenforced. More universally, decisions made while an agent is running escaped every reported system. With agent-generated pull requests rising sharply while human review capacity stays fixed, the next challenge is binding decisions across tools and sessions. One engineer turned a public debate about contradictory agent behavior into a promotion gate and CI check within a day and a half - evidence that safety rules can now emerge far faster than they once did.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.