A common safeguard for autonomous agents is to have a person approve their actions. But people tire, and a tired reviewer misses more. This paper asks how many actions should be sent to a human, and which ones, when the reviewer’s catch probability falls as the recent workload rises. We model a reviewer whose catch probability declines with recent busy time, and a risk scorer that ranks actions. We prove that, when the scorer’s ranking is well behaved, the share of bad actions caught is a concave function of the review rate, so there is a single best rate, and that for a Gaussian scorer reviewing everything is never the best choice once fatigue is present. At the optimum, an action is worth sending to the reviewer only if its probability of being bad exceeds a threshold that depends on the fatigue strength and on how much is already being reviewed. In a minute-by-minute simulation of an eight-hour shift (1,000 runs per setting), sending every action to the reviewer missed 17.3 of about 30 bad actions, because the mean catch probability fell to 0.43. Sending only the riskiest 36% of actions missed 11.2, a reduction of 35% with 67% fewer reviews. A threshold computed from the theory, with no tuning, matched the best tuned threshold. A controller that simply holds the reviewer’s load near a target did not beat a well-tuned fixed threshold, and learning the fatigue strength from audits worked only when about 30% of reviewed bad actions were audited. Whether scheduled breaks help depends on the fatigue model: they did nothing when fatigue followed recent load alone, and cut misses by 17 to 23% when a slow, cumulative component was added. All results are from simulation with assumed parameters.
