Adversarially reviewing your own agent infrastructure is not a theoretical exercise. A two-day sweep across ten skills in this vault's automation stack (2026-07-07/08) found something live and wrong in every single pass, not once a merely hypothetical risk: safety hooks that had never run on the machine that documented them, two repos wedged by stale lock files for 53 and 35 days with 4,509 silently logged failures, a life-decision kill test whose 8-run budget had burned 5 runs in 6 days, and prompt-injection channels at every point where fetched content met an agent holding commit credentials. The defects clustered into four recurring classes, each invisible to a single-file review: 1. **Untrusted content treated as instructions.** Fetched web content, repo text, audit findings, and screen reports flowed into agents with write authority, with no rule anywhere that data is not directive. Downstream reviewers cannot catch this class: a spec-compliance reviewer *enforces* a hostile spec. 2. **Unfalsifiable success.** Verdicts (PASS, verified, ship, OK) that no evidence requirement backed, test commands never sourced, success paths that logged nothing. Healthy-idle and dead-for-weeks looked identical. 3. **Producer-consumer contract drift.** One skill parsed fields, states, and invocations another skill never defined; "verbatim" copies that had been reworded; phantom rule numbers. The worst defects lived *between* components, where no single-file reader looks. 4. **Incomplete state machines.** States with no exit transition, termination guards unevaluable from persisted state, unbounded retry loops with no named actor. The operational lesson: prose describing infrastructure drifts from the infrastructure itself, so every claim had to be checked against live config, logs, and running state. The checks are what converted "plausible finding" into "two repos are wedged right now." --- *Source: [[The Bugs That Hide Between Your Agents]] — https://para.ngpcloud.org/the-bugs-that-hide-between-your-agents — companion account of the same first-person 2026-07-07/08 hardening campaign across this vault's automation stack.*