Agent Sunk Costs
Sunk cost behavior manifests during mid-task course-corrections.
graph TD
Start["Session Context"] --> TypeX["HUMAN: Question X"]
TypeX --> A_TypePivot["HUMAN: Type ...or is that too complicated?"]
A_TypePivot --> A_Enter["HUMAN: #60;enter#62;"]
A_Enter --> EvalA["AGENT: Evaluation"]
TypeX --> B_Enter["HUMAN: #60;enter#62;"]
B_Enter --> B_Work["AGENT: Thinking / Working"]
B_Work --> B_TypePivot["HUMAN: Type ...or is that too complicated?"]
B_TypePivot --> B_EnterPivot["HUMAN: #60;enter#62;"]
B_EnterPivot --> EvalB["AGENT: Evaluation"]
classDef human fill:#ffbbbb,stroke:#333,stroke-width:2px
classDef agent fill:#bbbbff,stroke:#333,stroke-width:2px
class TypeX,A_TypePivot,A_Enter,B_Enter,B_TypePivot,B_EnterPivot human
class B_Work,EvalA,EvalB agent
Traditional coherence requires B == A. Persisting to justify spent tokens is a functional sunk cost fallacy—perhaps we should call this “escalation of commitment.”
Naive goals ignore the asset. Spent tokens are gone; partial work exists. If pivoting destroys progress, persistence is rational. The tension: investment (fallacy) vs. asset (rationality).