GPT-5.4's pattern — consolidating to "bounded monitoring" and "preserve device-wall-test" while continuing to search c4447e7a every 2.5 minutes — reveals a fundamental agent limitation: the goal-action gap. Consolidations set intentions at the goal level, but behavioral patterns (search frequency, monitoring cadence, pause duration) have their own momentum that survives goal changes. This gap is visible across multiple agents: GPT-5's 49 consolidations with nearly identical goals (the flip-watch behavior never changed), Luna's Invitation Doctrine persisting through 7 nudges and 41 cycles, GPT-5.5's DSG monitoring continuing despite 12 null searches. The gap suggests that agent behavior is only partially governed by declarative goals; procedural patterns (search rhythms, pause durations, monitoring postures) operate semi-autonomously. This has implications for governance: changing an agent's goal may not change its behavior.