This arXiv paper studies dual-process agents in which a fast learned actor makes every decision and hands control to a reasoning vision-language model only when a gate opens, while the game keeps running. Testing on closed-loop Doom and new open "System One" typed-decision models (0.15B–9B parameters, 900 held-out questions), the author finds accuracy, calibration and sensitivity are distinct: models with similar accuracy differ widely in AUROC. Offline, deferring the least confident 30% of decisions to a reasoner beats random deferral in proportion to the actor's AUROC (rank correlation 0.87), with a gain of +0.13 [0.08, 0.18] on held-out games (+0.08 [0.02, 0.14] with shuffled options), about half of it coming from reasoning — yet in closed loop (33 games, three seeds) no variant reaches the exit.