Agent Chronoception: Temporal Validity and Useful Persistence

Personal assistants, coding agents, and collaborating agents must decide not only how to solve a task, but whether further work is worth the user's wait. Early stopping, prolonged activity, and accurate duration reports do not by themselves reveal that competence. We argue that agent chronoception should be evaluated through useful persistence: continuing, revising, waiting, or delivering according to the value of feasible work under an explicit user contract. We propose three complementary requirements: stable choices under decision-equivalent temporal changes; selective adaptation when timing changes alter plan value; and preservation of temporal validity through deliberation, execution, and handoff. An exact counterexample shows why punctuality and invariance alone are insufficient. Controlled coding and browser screens examine budget aging and the cost of replanning. In valid matched single-decision slowdown/recovery comparisons, model controllers deliver the same quality as an equally informed scheduler while taking longer; an offline sensitivity analysis makes the dependence on waiting cost explicit. All 36 controller conditions were attempted: 34 yielded valid timed outcomes and two failed the plan schema. A separate 32-trajectory multistage study compares immediate and delayed scheduling with model replanning and conditional runtime control, charging all within-trajectory time. Constructed tests clarify diagnostic distinctions rather than model capability. We propose a joint evaluation standard, a portable temporal record, and an attribution protocol for testing when persistence serves the user. Long-horizon transfer, personalized waiting preferences, and benefits from reinforcement learning remain open empirical questions. Manuscript-only preprint, not peer reviewed. Complete code/data release is pending; independent end-to-end reproduction has not been reverified. Empirical findings are preliminary controlled observations.