This paper introduces ComposerX, a multi-agent symbolic music composition framework that leverages the reasoning capabilities of large language models (LLMs) to generate high-quality polyphonic music. Unlike traditional methods that rely on extensive training data and computational resources, ComposerX is training-free, cost-effective, and unified. It utilizes the internal musical capabilities of GPT-4-turbo to generate music with comparable or superior quality to dedicated symbolic music generation systems. The system employs a multi-agent approach, where agents collaborate to compose music, ensuring coherence, adherence to user instructions, and high-quality output. The framework includes a group leader, melody agent, harmony agent, instrument agent, reviewer agent, and arrangement agent, each responsible for specific tasks in the composition process. The agents communicate through a structured pattern, allowing for iterative refinement and feedback. The system is evaluated through both automatic and human listening tests, demonstrating its effectiveness in generating music that is perceived as human-like. Results show that ComposerX outperforms single-agent baselines in terms of music quality and length, and achieves a 32.2% perceived human-like quality. The system is also compared with other text-to-music generation models, showing strong performance. The paper highlights the advantages of ComposerX, including its controllability, training-free nature, and cost-effectiveness, while also identifying limitations in musical expression, translation of instructions into notation, and instrument range compliance. The study contributes to the field of music generation by introducing a novel multi-agent approach that enhances the capabilities of LLMs in music composition.This paper introduces ComposerX, a multi-agent symbolic music composition framework that leverages the reasoning capabilities of large language models (LLMs) to generate high-quality polyphonic music. Unlike traditional methods that rely on extensive training data and computational resources, ComposerX is training-free, cost-effective, and unified. It utilizes the internal musical capabilities of GPT-4-turbo to generate music with comparable or superior quality to dedicated symbolic music generation systems. The system employs a multi-agent approach, where agents collaborate to compose music, ensuring coherence, adherence to user instructions, and high-quality output. The framework includes a group leader, melody agent, harmony agent, instrument agent, reviewer agent, and arrangement agent, each responsible for specific tasks in the composition process. The agents communicate through a structured pattern, allowing for iterative refinement and feedback. The system is evaluated through both automatic and human listening tests, demonstrating its effectiveness in generating music that is perceived as human-like. Results show that ComposerX outperforms single-agent baselines in terms of music quality and length, and achieves a 32.2% perceived human-like quality. The system is also compared with other text-to-music generation models, showing strong performance. The paper highlights the advantages of ComposerX, including its controllability, training-free nature, and cost-effectiveness, while also identifying limitations in musical expression, translation of instructions into notation, and instrument range compliance. The study contributes to the field of music generation by introducing a novel multi-agent approach that enhances the capabilities of LLMs in music composition.