AlphaCrafter is a full-stack multi-agent framework for cross-sectional quantitative trading that addresses the fragmentation problem in existing quantitative workflows. It unifies factor discovery, regime-aware selection, and risk-constrained execution within a single closed-loop system that adapts continuously to evolving market conditions without manual intervention. The framework consists of three specialized agents: a Miner that expands the factor pool via LLM-guided search with validation, a Screener that assesses market regimes to construct diversified factor ensembles, and a Trader that optimizes and executes strategies under explicit risk constraints. Together, they form a hypothesis–validation–execution loop that dynamically reconfigures signal ensembles and portfolio construction as market regimes shift.

Extensive experiments on CSI 300 and S&P 500 markets demonstrate that AlphaCrafter consistently outperforms state-of-the-art baselines—including traditional quantitative methods (MACD, Grid Trading), machine learning (LightGBM, XGBoost), deep learning (LSTM, Transformer, TRA), and LLM-based agents (TradingAgents, TradingGroup, RD-Agent, AlphaAgent)—in risk-adjusted returns while exhibiting the lowest cross-trial variance. Notably, AlphaCrafter is the only method that delivers consistently positive risk-adjusted returns across both backtesting and live trading phases, avoiding the catastrophic losses seen in all baselines. Ablation studies confirm the marginal contribution of each agent, and alpha decay analysis shows that dynamic factor curation is essential for maintaining predictive power. The framework also demonstrates robustness to the choice of underlying LLM, with stable performance across GPT, Claude, and Gemini backbones.