Language‑model agents can now perform long‑duration scientific tasks in biology, yet their judgments about conventions, skills, errors, evidence, and stopping are embedded within the reasoning process, making them hard to measure, replace, or improve. Operon‑S1 introduces a fast decision system (System 1) that separates judging from doing. At 13 decision points the system poses the agent’s recurring judgments as closed‑ended questions, sending only fixed messages that never contain an answer; any model that can answer these questions serves as the judge.\n\nAcross seven life‑science benchmarks, a local decision model answered these questions with a median latency of 0.6 s on CPU and caused no measurable change in the agent’s score on five of six benchmarks, whereas the general language model driving the agent consumed 8.0%–35.2% of task cost for the same queries. Every decision is logged together with its consequence, and retraining on labels derived from task construction, verifiers, or execution improves accuracy on every question for all three trained versions; a new version is released only after passing a promotion gate.\n\nThe same closed decisions are applied to wet‑lab protocol troubleshooting and the design of a germline target enrichment panel. Like an operon regulator that switches genes on and off without producing the gene products, Operon‑S1 steers the agent without doing its work, turning the agent’s judgments into a measurable, replaceable, and trainable layer.\n\nReview