Code agents produce artifacts—files, deployments, database schemas—that persist after the context window moves on. The agent forgets; the artifact does not. During a development session with Claude Code, we observed five failures rooted in this asymmetry: artifact amnesia (F1), false positive verification (F2), state confabulation (F3), temporal schema incoherence (F4), and detection exclusively by the human operator (F5). The context window is not just a length limit—it is an epistemic boundary beyond which the agent cannot distinguish what it verified from what it assumed, what it built from what it found. Second paper in a series; the first addressed LLM-as-judge non-determinism.
H. Tamba (Tue,) studied this question.