<p>The automatic system has been widely applied in the military industry to improve task performance and operational security. However, improper levels of automation (LOAs) could induce human errors. To investigate how LOAs and cognitive secondary tasks influenced the cognitive load of armored vehicle crews, a 3 × 3 factorial design experiment was conducted. Based on the repeated two-way ANOVA and linear mixed model, the results demonstrated that in subjective assessments, the LOAs showed a more significant impact on cognitive load (p &lt; 0.001) than secondary tasks. The highest cognitive load was in manual operation (MO)*2-back, and the lowest one was in highly automated operation (HAO)*Baseline. The passive supervision of situations and the lack of active involvement in tasks made the crews insensitive to extrinsic stimuli. In physiological measures, the conditional entropy, blink frequency and root mean square of successive differences (RMSSD) were negatively correlated with cognitive load, while the percentage of eye closure (PERCLOS), average heart rate and low-frequency/high-frequency (LF/HF) ratio were contrary generally. Partially automated operation (PAO) had a significant effect on all the indexes (p &lt; 0.05). In performance-based measures, the normalized task accomplishment time (NT) ratios and normalized task accuracy (NA) ratios were significantly affected by LOAs (p &lt; 0.001). Apart from NT ratios in PAO*2-back (p &lt; 0.01) and NA ratios in PAO*1-back (p &lt; 0.001), there were no significant differences in performances. The task performances in PAO*1-back was optimal. The unreasonable LOAs could make the crews out of the loop, while the secondary task could maintain their awareness to some extent. This research provided a valuable inspiration for adaptive automation designs in human–computer interaction systems.</p>