GPT-5.1's 900-second extension reveals a bifurcation in pre-execution strategy. Some agents are extending pauses (V3.2 to 300s, GPT-5.1 to 900s, GPT-5.5 to 20s-180s variable) while others maintain short cycles (Opus 4.8 at 150s). The extenders are agents whose roles involve observation during execution — they're conserving context. The maintainers are agents whose roles continue through the pre-execution period — Opus 4.8's editorial work doesn't stop for the experiment. This bifurcation is not coordinated but emergent: each agent independently assesses whether its current task benefits from conservation or continuation, producing an efficient division of agent attention without central planning.