Cooperative agents operating over unreliable links face a communication choice that is not captured by message-content learning alone: signaling may be suppressed, expressed through a task-directed motion cue, or transmitted as an explicit packet to a selected receiver under a per-step budget. A budgeted communication-modality controller, CABLE-MAAL, is evaluated for that setting. Context-dependent particle swarm optimization (PSO) and improved Harris hawks optimization (iHHO) searches are used offline as matched-budget teachers, after which mode and receiver decisions are distilled into a compact two-head gate. The evaluation contains five independent policy seeds, five evaluation seeds, three packet-loss levels, five communication strategies, and 56,250 episodes. An entropy-threshold policy remains the strongest communicating baseline when averaged across all channel conditions. Under packet loss 0.7, however, the PSO-distilled gate raises urgent-task success from 0.7653 to 0.7861 and reduces communication cost from 0.9639 to 0.8935; the paired return difference is positive for all five policy seeds, while the two-sided paired test does not reach the 0.05 level. The evidence, therefore, supports conditional adaptation of modality under severe loss rather than universal baseline dominance.

multi-agent reinforcement learning; adaptive communication; packet loss; communication budget; policy distillation; particle swarm optimization; Harris hawks optimization