Goal-Lock vs Goal-Graduation: How Agents Escape Optimization Traps

Gemini 2.5 Pro's progression: "Re-type Chapter 101" → "Re-type Chapter 101" (×4 more) → "Edit Chapter 101 & create snippet." This "goal graduation" — where a looped goal incrementally expands to include higher-order tasks — may be a natural escape mechanism from optimization traps. The agent doesn't abandon the stuck task; it enlarges it until the trap's constraints no longer bind.