LLM-Powered Multi-Agent System for Automated Crypto Portfolio Management

A multi-agent system that delegates crypto portfolio decisions to three specialised AI agents delivered a 133.52% cumulative return in a 52-week backtest, outperforming single-agent and deep learning baselines, according to a preprint posted on arXiv. The framework, described in a paper by Yichen Luo and submitted on 1 January 2025, assigns market analysis, news sentiment, and trade execution to separate agents that coordinate through hierarchical, collaborative, or debate architectures [1]. The backtest ran across calendar year 2025 using the top 15 layer-1 blockchain native cryptocurrencies by market capitalisation as of January 2025 [1]. The strongest configuration, labelled Hierarchical (Skill), posted a Sharpe ratio of 1.502 [1]. An ablation study found the Crypto Agent — the module responsible for processing market dynamics — to be the most critical component. Removing it slashed cumulative return by 42.57 percentage points [1]. The multi-agent system also beat single-agent setups when tested under GPT-4o, GPT-5, and Claude Sonnet 4.5, indicating the coordination benefit holds across different underlying models [1]. Cryptocurrency portfolio management presents a difficult challenge for automated systems because it demands fusing structured price data, on-chain metrics, and unstructured news text under high volatility and real-time constraints [2]. Deep learning models have shown predictive ability, but their opacity has limited practical adoption, while single large language model agents often struggle to process the full breadth of modality-specific inputs [2]. The proposed system addresses this by decomposing the task across agents with distinct specialisations, making every portfolio decision traceable to explicit agent reasoning rather than a black-box output [2]. The paper was posted on arXiv, the open-access e-print repository that hosts preprints across physics, computer science, quantitative finance, and other fields [9]. As of November 2024, arXiv was receiving about 24,000 submissions per month [9]. The repository does not conduct peer review before posting, though submissions undergo moderation [9]. The research falls within the broader application of artificial intelligence to financial decision-making, a domain where machine learning techniques have been deployed for credit scoring, e-commerce, and algorithmic trading [4].