The ACLArena framework was introduced to investigate and optimize Agent Continual Learning (ACL) during multi-stage post-training, addressing the challenge of integrating diverse capabilities without catastrophic forgetting. The researchers established a sequential training pipeline to analyze forgetting and generalization mechanisms through both model-level and token-level perspectives. By systematically evaluating integration paradigms—including multi-teacher on-policy distillation, self-distilled fine-tuning, and model merging—the study provides a detailed understanding of how capabilities transfer across training stages. To mitigate performance degradation, the authors proposed a novel ACL recipe that utilizes offline replay of high-quality trajectories combined with a routed network of multiple LoRA experts, where each expert is specialized via Reinforcement Learning (RL). Experimental results across four distinct reasoning and agentic tasks, evaluated in both in-domain and out-of-domain environments, demonstrate that this approach significantly enhances an agent's ability to acquire new skills while preserving previously learned ones. This research provides a comprehensive evaluation framework and a practical methodology for deploying general-purpose agents in complex industrial settings through effective multi-stage capability integration.