AI agents in finance: what three 2025 papers actually do

Three papers get passed around whenever "AI agents in finance" comes up. This page reads each one from its arXiv abstract and, where there is one, its repository README, and says what the authors claim and nothing beyond it. It is not trading or investment advice, and Bindler does not build or sell trading tools: for an unlicensed company those are regulated activity. What is worth taking from these papers is a working pattern, and that is where the page ends.

1. TradingAgents (Xiao, Sun, Luo, Wang, arXiv 2412.20138, December 2024)

What it is. A multi-agent framework that copies the division of labour in a trading firm. The abstract lists the roles: fundamental, sentiment and technical analysts, Bull and Bear researcher agents that assess market conditions, a risk management team monitoring exposure, and traders who "synthesize insights from debates and historical data" to decide. What the code does. The README (TauricResearch/TradingAgents, v0.2.5 in May 2026) describes the pipeline in four teams. An Analyst Team (fundamentals, sentiment, news, technical) writes reports. A Researcher Team of bullish and bearish researchers "critically assess the insights provided by the Analyst Team" through structured debate. A Trader agent composes those into a decision on timing and size. A Risk Management team reviews it and a Portfolio Manager approves or rejects the proposal; an approved order goes to a simulated exchange. Every completed run is appended to a decision log, and on the next run for the same ticker the system fetches the realised return and writes a one-paragraph reflection. A backtest command runs the pipeline over a grid of tickers and dates and scores the decisions whose holding window has since traded. What the authors claim. Improvements over baseline models in cumulative return, Sharpe ratio and maximum drawdown in their experiments. The README states the framework "is designed for research purposes" and is not intended as financial, investment or trading advice. What to notice. The interesting part is not the trading. It is that the decision is produced by a debate with named sides, then a separate reviewer, then a log that is re-read against what actually happened.

2. HARLF (Coriat and Benhamou, arXiv 2507.18560, July 2025)

What it is. A hierarchical portfolio optimisation framework. Three tiers: base reinforcement learning agents process "hybrid data" (market indicators plus sentiment signals from financial news scored by lightweight language models), meta-agents aggregate the base agents' decisions, and a super-agent merges those on market data and sentiment. What the authors claim. Trained on 2000 to 2017 and evaluated on 2018 to 2024, the abstract reports a 26% annualised return and a Sharpe ratio of 1.2, against equal-weighted and S&P 500 benchmarks. Contributions named: cross-modal integration, a hierarchical structure for stability, and open-source reproducibility. What to notice. The language model's job here is narrow: turn news into a sentiment number that feeds a conventional learner. The "agent" is the reinforcement learner, not the LLM. A single backtest window is one sample; the abstract makes no claim about live trading.

3. Automate Strategy Finding with LLM in Quant Investment (Kou et al., arXiv 2409.06289, September 2024, revised 2025)

What it is. A three-stage framework. Prompt-engineered language models generate candidate "alpha factors" as executable code across diverse financial data; a multi-agent evaluation filters them on market status and predictive quality while keeping the categories balanced; a weighting step adapts the surviving factors to market conditions. What the authors claim. Robust performance in Chinese and US market regimes against established benchmarks, including a 53.17% cumulative return on the SSE50 from January 2023 to January 2024. The public repository under the first author's name carries the paper title and little else, so the claims here rest on the paper. What to notice. The LLM is a generator of hypotheses, and everything downstream exists to throw most of them away. The filter is the product.

The pattern that transfers

Strip the markets out and the three papers share one design: many narrow proposers, an adversarial or hierarchical review, one decision, and a record that is checked later against the outcome. That pattern does not need a trading desk. Bindler runs it on ordinary company decisions: departments propose, a judgement bench steelmans both sides and rules or sends the question to the founder, every ruling is written down with what would reverse it, and the daily runs report a miss rather than hide it. How that is set up, with the department files and the rules, is in How Bindler runs and the playbook.

What none of the three papers demonstrates is that the pattern makes money for a small company that buys it. The evidence is backtests on the authors' own data. Read them for the architecture, not the returns.

Sources

• Xiao, Sun, Luo, Wang, "TradingAgents: Multi-Agents LLM Financial Trading Framework", arXiv 2412.20138, 28 December 2024; repository github.com/TauricResearch/TradingAgents, README as of v0.2.5 (May 2026).
• Coriat, Benhamou, "HARLF: Hierarchical Reinforcement Learning and Lightweight LLM-Driven Sentiment Integration for Financial Portfolio Optimization", arXiv 2507.18560, 24 July 2025.
• Kou et al., "Automate Strategy Finding with LLM in Quant Investment", arXiv 2409.06289, 10 September 2024; repository github.com/kouzhizhuo/Automate-Strategy-Finding-with-LLM-in-Quant-investment.

Read on 20 September 2026 from the arXiv abstracts and the READMEs as they stood that day.