This paper examines the construction of an optimal execution policy using reinforcement learning, with the objective of identifying the most efficient strategy for executing buy and sell orders over a fixed trading horizon. The proposed framework incorporates high-frequency decision-making based on features extracted from the prevailing state of the limit order book, allowing for fine-grained control of execution dynamics. To model the trading environment and address the constraints inherent in historical market data, we employ the multi-agent market simulator ABIDES, which enables access to limit order books with varying levels of depth. We introduce a tailored Markov Decision Process (MDP) formulation and report the outcomes of the proposed approach, comparing its performance against established execution benchmarks. The findings indicate that the reinforcement learning agent consistently surpasses conventional strategies, highlighting its potential applicability in real-world trading settings.