Agile flight in cluttered environments presents unique control challenges, particularly for morphing drones subject to complex aerodynamic interactions. While Reinforcement Learning (RL) has emerged as a powerful tool for generating agile control policies, direct actuator control via RL often suffers from a lack of robustness due to model discrepancies, wind disturbances, and ground effects. This work addresses these limitations by investigating robust control strategies that augment or stabilize RL-based policies. Drawing on techniques proven in quadrotors and fighter aircrafts, we explore the implementation of Model Reference Adaptive Control (MRAC) methods, specifically L1-Adaptive Control (L1AC), to compensate for model discrepancies in real time.