Abstract
Language models have given rise to agentic AI: autonomous systems that pursue high-level goals through tool use, orchestration, and adaptation. Although prior surveys treat financial applications, agent architectures, and fine-tuning methods in isolation, this review integrates the full agentic stack. Covering literature from 2024 to 2026, we examine the structure of agentic systems, including orchestration topologies, MCP and A2A protocols, reasoning frameworks, and reinforcement-learning paradigms from RLHF to GRPO, together with their financial applications and risks. We also complement this analysis with a competitor review of two industry frameworks. Our synthesis suggests that hallucination and reproducibility are among the main constraints on autonomy, often more limiting in practice than architectural constraints. Human-in-the loop verification therefore remains necessary in current financial deployments.
Authors
