Can an agent's own recovery logic make a working run fail?

Yes. A causal study of recovery, a harness re-querying the environment after a stale or missing observation, measured the same task with and without that step and found it cuts both ways: on 300 long-horizon ALFWorld episodes, a policy built to decide when recovery was worth it still produced 2 cases where intervening broke a trajectory that would otherwise have succeeded, against 11 it rescued. An agent stuck in a loop with nothing ever throwing an error is the visible failure of too much recovery; this is the quieter one, where recovery that looks safe on average still carries a real harm rate that an overall success-rate number hides. The paper’s router, which only intervenes when the evidence available before the retry favors it, raised success from 70.33% to 73.33%, a smaller net gain than either count alone suggests, because the number that matters is rescues minus harms, not rescues by themselves. A recovery step that helps most of the time is still a step worth checking for the trajectories it quietly breaks.