Neural circuits rely on excitatory–inhibitory (E/I) interactions to support adaptive learning and decision-making. Here, we investigate how these dynamics contribute to flexible behaviour across three modelling levels. First, using a simplified mean-field model of two-choice decision-making, we examine the computational role of selective excitation and inhibition in stabilizing or amplifying competition between alternative choices. Building on these insights, we embed a similar E/I mechanism into the preference function of a reinforcement learning (RL) agent, showing how inhibitory feedback modulates behavioural adaptation in reversal learning. Finally, we assess the scalability of these principles by training RL agents with E/I-constrained recurrent neural networks (RNNs) in dynamic tasks. While a general E/I architecture allows broader forms of inhibitory influence, our results indicate it hinders learning in these settings. In contrast, a structured architecture enforcing local inhibition preserves biological plausibility while maintaining robust performance. Together, these findings suggest that E/I dynamics may provide a feasible computational mechanism for adaptive learning and decision-making.