When Assurance Work Displaces the Task: A Retrospective Case Study of Overdefensiveness and Overengineering in AI Agent Use

AI agents can impede work by adding confirmation, qualification, verification, and infrastructure, even when those activities appear individually reasonable. This paper examines how such work becomes disproportionate in one participant's sustained use of Codex. A retrospective, AI-assisted analysis revisits 32 episodes from 20 conversations selected from local records spanning April-November 2026, and adds three contrast episodes from three other conversations. The unit of analysis is a task-and-correction episode, not a model run or an independent failure. Overdefensiveness concerns risk handling beyond the demands of the task and its actual authorization; overengineering concerns complexity without a sufficiently demonstrated use. The trajectories show that process outputs can displace requested deliverables, extensive checking can miss the actual operating condition, and attempts to correct drift can themselves create further work. They also show that users sometimes authorize substantial assurance work and later reconsider it. Re-reading those authorizations changes the interpretation of apparent excess. The clearest local improvements involve recognizing existing capabilities, removing low-value supporting material, and aligning acceptance with the actual use path. Necessary safeguards and required independent review remain legitimate. The contribution is a task-relative account of assurance work, a deidentified episode catalogue, and propositions for subsequent evaluation. The evidence does not estimate prevalence, establish a treatment effect of personal instructions, or identify a model or harness as the principal cause. This record contains the English research preprint (17 pages), a deidentified catalogue of 35 episodes from 23 conversations, and a codebook. The author is also the participant; Codex assisted analysis and writing. Raw conversations and the identifying source crosswalk are not included. Version 1.0; not peer reviewed.