Agentic Financial Trading Agents:
A Comprehensive Literature and Research Survey

This paper presents the first comprehensive survey of agentic AI systems in financial trading, covering over 60 papers, frameworks, and repositories published between 2021 and 2026. The survey was conducted across Semantic Scholar, arXiv, GitHub, YouTube, and commercial product research.

Key Findings

1. Five Paradigms Identified
The field breaks into five distinct approaches: pure deep reinforcement learning (DRL) agents, LLM-augmented DRL hybrids, pure LLM-based trading agents, multi-agent collaborative systems, and market simulation environments. Each has different trade-offs in performance, interpretability, and deployment readiness.

2. Open-Source Dominates Innovation
Unlike most AI domains where commercial products lead, open-source frameworks dominate agentic trading. The FinRL ecosystem (15,308+ GitHub stars) and TradingAgents (81,614+ stars) far surpass any commercial offering. No major broker or FinTech platform currently offers autonomous LLM-based agency at the level of open-source research frameworks.

3. Hybrid LLM-RL Shows Most Promise
FLAG-Trader (25 citations, ACL 2025) and FinRL-DeepSeek (18 citations) demonstrate that combining LLM-processed financial text with RL training significantly outperforms either approach alone. However, true joint end-to-end optimization remains unsolved.

4. Multi-Agent Architectures Replicate Trading Firms
The most architecturally sophisticated systems (TradingAgents, HedgeAgents, FinVision) decompose trading into specialized roles — analysts, risk managers, portfolio managers — with structured communication protocols mimicking professional trading firms.

5. Ten Critical Research Gaps
The survey identifies ten gaps: unified LLM-RL optimization, regime change robustness, realistic market impact, standardized risk-aware metrics, behavioral consistency of LLM agents, regulatory frameworks, security/hallucination risks, real-time latency, cross-market generalization, and natural-language-driven strategy configuration.

Methodology

The review draws from systematic searches of Semantic Scholar, arXiv, GitHub, and web sources. Key papers were assessed by citation count, venue prestige, and GitHub popularity. The full paper includes a taxonomy table comparing approaches across data modalities, memory architecture, learning paradigm, risk management, and deployment mode.

Download the full paper from ResearchGate: