Understanding how humans perceive and adapt to artificial agents in competitive, multi-agent settings is essential for designing socially and strategically competent robots. We conducted two experiments using Chef’s Hat, a four-player card game, to examine how agent strategy and embodiment influence human perception, and adaptation. In both experiments, participants played against a mix of random and Deep Q-Learning (DQL) agents, one of which was embodied by a humanoid robot (iCub). In the first experiment, the robot followed a random policy, while the DQL agent was embodied by a box; in the second, the robot embodied the DQL strategy. Across both experiments, human players consistently outperformed the agents. DQL agents exhibited strategic preferences, such as discarding mid-value cards, leading to better performance than random agents, especially when controlling for game position. Despite similar final scores, participants often viewed the DQL agent as more intelligent, although the difference was not statistically significant. Embodiment alone did not significantly influence these perceptions, as participants focused more on gameplay dynamics than on the robot itself. However, when paired with a strategic policy, the robot’s embodiment amplified the recognition of its intelligence. These findings suggest that in competitive environments, the effectiveness of embodied agents depends not only on their strategies but also on gameplay context and player expectations, shaping how human players interpret their capabilities.