An agent takes 50 steps to resolve an incident. During execution, the full reasoning chain exists in the context window — why it chose path A over B, what evidence it weighed, which hypotheses it discarded. The moment the session ends, the output (the resolution) survives but the decision chain evaporates permanently. You know the agent fixed the issue; you cannot reconstruct why it chose that fix over alternatives.
This is distinct from the general context limitation described by The context window is the fundamental constraint — everything else follows. That insight addresses context as a scarce resource requiring active management. Reasoning evaporation names a specific, irreversible failure mode: the permanent loss of multi-step reasoning chains that existed only ephemerally in the window. The fix is not better context management but architectural — capturing traces during execution, not after. This is exactly what Decision traces are the missing data layer — a trillion-dollar gap proposes at the enterprise level: making “why” a first-class data type. And Session capture turns ephemeral AI conversations into a compounding knowledge base demonstrates the lightweight version — shadcn’s /done pattern captures decisions before they evaporate, turning ephemeral sessions into persistent knowledge. The gap between “most teams capture logs” and “traces capture reasoning” is the difference between knowing something happened and knowing why — and that gap widens with every agent session that closes without a trace.