Tool-using agents fail in recurring and classifiable ways, and the choice of recovery strategy is increasingly treated as something to be learned rather than fixed in advance. A designer weighing a hand-written recovery table against an online learner needs to know at what deployment horizon the learner overtakes the table. Existing crossover analyses assume the policy observes the true failure type, which real systems do not provide; the policy sees a classifier’s prediction. The present work reformulates the crossover problem under an arbitrary confusion matrix by mapping the strategy-success matrix into predicted-type space with an exact Bayes mixture, then separates the resulting degradation into a reward-ceiling channel and a learnability channel. At a classifier with 0.939 mean diagonal accuracy, ceiling loss exceeds learnability loss by roughly six to one. The two channels also work against each other: classifier noise raises the static table’s regret rate by 58%, yet shifts the predicted crossover only from 3053 to 3126 episodes, and a constructed instance establishes that the horizon is not monotonic in classifier accuracy. The reformulation’s free constant was tested under a protocol registered before the supporting data existed. The held-out ratio R = 1.4729 clears its pre-committed threshold of 1.5, but the bootstrap interval [1.00, 9.55] required by the same protocol does not confirm that the clearance is stable, so the component is reported as partially supported. A labeled, non-pre-registered follow-up traces the instability to one confound: all 25 environments were scored against a single reused table rather than one competent for each. Widening to 150 environments with a table authored per environment cuts median error from 59.7%/40.5% to 19.98%/14.00%, brings R to 1.4270, and narrows the interval to [0.78, 2.17], which still crosses 1.5, so the corrected result counts as substantially more stable rather than confirmed. Index Terms—autonomous agents, failure recovery, multiarmed bandits, Thompson sampling, noisy context, preregistration, negative results

autonomous agents, failure recovery, multi-armed bandits, Thompson sampling, noisy context, pre-registration, negative results