Project Bathtub: A Reliability-Engineering Approach to Failure-Mode Taxonomy for AI Coding Agents

Autonomous coding agents fail in ways that are numerous, unevenly costly, and unevenly hard to detect, yet most public discussion of agent reliability still centers on a single failure family: hallucinated APIs and factual errors. This report introduces Project Bathtub, an effort to import a decades-old industrial reliability discipline — Failure Mode and Effects Analysis (FMEA) — into the domain of LLM-based coding agents. We present Phase 0 of the project: an 80-item failure-mode taxonomy spanning ten categories, each scored on Severity, Occurrence, and Detectability (1–10). Ranking by RPN surfaces a finding at odds with the prevailing narrative: verification and self-assessment failures and specification/communication failures score as high or higher than any hallucination-related failure mode.