Agentic AI for Quantitative Finance: A Decision Science Framework and Scoping Review of Intelligent Alpha Systems

Alpha generation is increasingly shaped by large language models (LLMs) and agentic AI, yet much of the literature still treats alpha primarily as a predictive output rather than the result of sequential investment decisions under constraints. This paper presents a PRISMA-ScR-aligned scoping review and evidence synthesis of 165 studies published from 2023–2026. The review examines how LLM-assisted and agentic approaches contribute to forecasting, information interpretation, portfolio construction, execution, and coordination. The evidence shows that LLM-assisted systems substantially outnumber agentic systems. Predictive, interpretive, and execution-oriented applications are represented at broadly similar levels, while coordination-oriented systems form a smaller but distinct class. Validation remains dominated by historical and backtest-based evidence, with live or forward evaluation comparatively rare. Building on these findings, the paper proposes a five-stage decision capability spectrum, an evidence-guided alpha typology distinguishing predictive, interpretive, execution, and coordination roles, and a decision-centric evaluation framework spanning performance, decision quality, robustness, explainability, and governance. Together, these contributions provide a structured basis for evaluating adaptive, auditable, and constraint-aware intelligent alpha systems.