GuruAgents: Prompt-Guided Investment Agents
- GuruAgents are a family of prompt-guided, tool-augmented LLM agents that emulate legendary investors by converting qualitative doctrines into explicit, rule-based portfolio strategies.
- They integrate a GPT-4o backbone, investor-specific system prompts, and rigorous financial tools that preprocess and score market data through a deterministic reasoning pipeline.
- Backtests show varied performance—with the Buffett-inspired agent achieving the highest CAGR—illustrating both the proof-of-concept nature and future potential for broader financial applications.
Searching arXiv for papers directly relevant to GuruAgents and closely related agent frameworks. Searching for "GuruAgents prompt-guided LLM agents" and related terms on arXiv. GuruAgents denotes a family of prompt-guided, tool-augmented LLM agents introduced to emulate the investment styles of specific “guru” investors—Benjamin Graham, Warren Buffett, Joel Greenblatt, Joseph Piotroski, and Edward Altman. In the original formulation, each GuruAgent uses GPT-4o, a guru-specific system prompt, financial metric tools, and a deterministic reasoning pipeline that converts structured fundamentals and market data into stock scores, portfolio weights, and short rationales (Kim et al., 2 Oct 2025).
1. Concept and historical positioning
GuruAgents were proposed to address a specific translation problem in quantitative investing: legendary investors typically articulate strategy in qualitative, principle-driven terms—such as margin of safety, moats, distress zones, Magic Formula, and F-Score—whereas systematic portfolio construction requires explicit, reproducible rules. The framework asks whether a properly prompted LLM, equipped with financial tools, can operationalize those doctrines without reducing them to generic factor models (Kim et al., 2 Oct 2025).
In this original sense, GuruAgents are not black-box return predictors. They are rule-following agents whose behavior is constrained by system prompts that specify role, creed, available tools, and explicit scoring and portfolio-construction procedures. The LLM is therefore used as a controller for a quant-style workflow rather than as an unconstrained forecasting model (Kim et al., 2 Oct 2025).
The five implemented personas are summarized below.
| GuruAgent | Philosophy | Core scoring basis |
|---|---|---|
| Benjamin Graham | Value investing, margin of safety, financial strength | Weighted liquidity, profitability, and efficiency with leverage penalties |
| Edward Altman | Distress screening via Z-Score | Normalized , , or with safe/grey/distress cutoffs |
| Joel Greenblatt | Magic Formula | Combined rank of earnings yield and ROIC |
| Joseph Piotroski | Accounting-based quality improvement | -Score out of 9 |
| Warren Buffett | “Wonderful companies at fair prices” | Quality, coverage, profitability, valuation, and cash-based bonuses/penalties |
2. Prompt architecture, tools, and deterministic workflow
Each GuruAgent is built from four components: a GPT-4o backbone, a system prompt encoding a specific investor’s philosophy, a set of financial tools, and a deterministic reasoning pipeline (Kim et al., 2 Oct 2025). The system prompts have three parts. First, they define role and creed through investor-specific language and canonical quotes. Second, they specify the data schema and the available tools. Third, they prescribe an explicit scoring and portfolio-construction procedure.
The workflow is quarterly. The agent receives NASDAQ-100 constituent data, including quarterly fundamentals and quarter-end market snapshots. Some agents also use trailing-twelve-month and year-over-year quantities constructed through tools. The prompt instructs the agent to call the relevant metric functions once each on the full DataFrame, then perform all subsequent ranking and weighting from tool outputs alone (Kim et al., 2 Oct 2025).
Preprocessing is standardized across agents. Metrics are winsorized, typically at 5–95%, then min–max scaled to ; if a metric has no variance, it is assigned $0.50$. For “lower is better” variables such as valuation multiples, the prompt specifies inversion through . Agent-specific formulas then generate a base score, after which bonuses and penalties are applied, clipped to , tie-breakers are resolved deterministically, and weights are normalized. The general portfolio rule is:
with final weights rounded to integer percentages and the last line absorbing rounding error (Kim et al., 2 Oct 2025).
The implementation uses LangChain and LangGraph, with tool functions defined outside the LLM. This arrangement preserves determinism at the level of metric computation while allowing the prompt to express doctrine, ranking logic, and short stock-level rationales (Kim et al., 2 Oct 2025).
3. Investor personas and scoring logics
The Graham GuruAgent is framed around balance-sheet strength, low leverage, and simple testable rules. Its toolset includes current ratio, debt-to-equity, interest coverage, ROE, asset turnover, profit margin, working capital ratio, and valuation. Its score is a weighted combination of liquidity, profitability, and efficiency, with explicit penalties for high leverage or weak coverage and bonuses for strong working capital and current ratio (Kim et al., 2 Oct 2025).
The Altman GuruAgent centers on bankruptcy-risk classification through the Z-Score family. Its prompt specifies which formula variant to use and the corresponding safe, grey, and distress cutoffs. Score normalization is performed between those cutoffs, and selection prioritizes safe-zone firms, adding grey-zone names if necessary to reach the portfolio size rule (Kim et al., 2 Oct 2025).
The Greenblatt GuruAgent encodes the Magic Formula. It uses tools for earnings yield, ROIC, safety, and size/liquidity, excludes firms with negative EBIT or nonsensical denominators, ranks eligible firms by earnings yield and ROIC, and converts the combined rank into a score. Safety nudges are comparatively mild, which preserves the formula’s mechanistic character (Kim et al., 2 Oct 2025).
The Piotroski GuruAgent implements the nine binary F-Score signals covering profitability, leverage/liquidity/funding, and operating efficiency. Missing values count as zero, eligibility requires at least four evaluable signals, and ranking proceeds by F-Score with tie-breakers on contemporaneous profitability and margin improvement. This agent is therefore strongly driven by changes in accounting quality rather than by valuation alone (Kim et al., 2 Oct 2025).
The Buffett GuruAgent is the most elaborate. In addition to debt-to-equity, interest coverage, ROE, profit margin, asset turnover, and valuation, it uses FCF yield and ROCE, and the prompt further defines derived components such as CashConversion, MarginStability, BuybackYield, CapExIntensity, and OwnerEarningsYield. Its final score is the sum of a weighted base, a QualityPlus component, and explicit bonuses and penalties. The construction systematically favors large, profitable firms with strong cash flows, conservative balance sheets, and non-extreme multiples (Kim et al., 2 Oct 2025).
4. Backtest results, portfolio behavior, and benchmark comparison
The reported backtest covers NASDAQ-100 constituents from Q4 2023 to Q2 2025. The Buffett GuruAgent delivers the highest CAGR, approximately 42.23%, with an annualized Sharpe of about 1.40 and maximum drawdown of about -22.34%. Piotroski is second in CAGR at about 30.93%, followed by Graham at about 28.74%, Altman at about 25.74%, and Greenblatt at about 19.38% (Kim et al., 2 Oct 2025).
Against benchmarks, NASDAQ-100 posts a CAGR of about 29.36% and annualized Sharpe of about 1.32, while the S&P 500 posts a CAGR of about 26.31% and annualized Sharpe of about 1.47. Thus Buffett achieves the highest absolute return among the compared strategies in this window, whereas the S&P 500 remains slightly higher on annualized Sharpe (Kim et al., 2 Oct 2025).
The agents also exhibit distinct behavioral profiles. Buffett produces concentrated portfolios with repeated exposure to AAPL, MSFT, and NVDA and lower turnover, consistent with the “hold for the long term” prompt. Piotroski shows high turnover because the F-Score is re-evaluated frequently from changing fundamentals. Greenblatt sits in between, reflecting quarterly reranking by formula. Altman tilts toward financially safe firms, while Graham emphasizes broad financial strength rather than franchise quality (Kim et al., 2 Oct 2025).
The paper does not report formal statistical significance tests, granular ablations that remove individual prompt components, or long-horizon robustness studies. Its main empirical claim is therefore comparative and descriptive: different prompted doctrines, when paired with the same base LLM and comparable data, yield materially different portfolios and performance profiles (Kim et al., 2 Oct 2025).
5. Broader architectural meaning in later agent literature
Although GuruAgents originally denotes an investment framework, later papers use “GuruAgents-like” to name a broader design pattern: specialized, tool-augmented, auditable agents with explicit observation, reasoning, and action channels. In web automation, "EmbeWebAgent" argues for embedding agents directly into controlled enterprise UIs through curated ARIA labels, URL-based observations, a per-page function registry, and a backend workflow, explicitly identifying this as relevant to a “GuruAgents”-style framework (Ma et al., 16 Feb 2026).
In software engineering, "Knowledge-Guided Multi-Agent Framework for Application-Level Software Code Generation" presents a COPA–CA–TA architecture in which architect, coding, and testing agents operate in a closed loop over SRS and ADD artifacts; its discussion explicitly maps these design choices to a GuruAgents-like system for code generation (Xiong et al., 22 Oct 2025). In evaluation, "What Is Your Agent's GPA? A Framework for Evaluating Agent Goal-Plan-Action Alignment" proposes Goal Fulfillment, Logical Consistency, Execution Efficiency, Plan Quality, and Plan Adherence as reference-free LLM-judge metrics for agent traces, which is directly relevant to any GuruAgents-style orchestration layer (Jia et al., 9 Oct 2025).
More recent orchestration work pushes this abstraction further. "AOrchestra: Automating Sub-Agent Creation for Agentic Orchestration" models any agent as a tuple 0 of Instruction, Context, Tools, and Model, then delegates via on-the-fly sub-agent creation (Ruan et al., 3 Feb 2026). "MetaAgent: Toward Self-Evolving Agent via Tool Meta-Learning" shows a frozen-weight agent improving through help-seeking, tool routing, self-reflection, and a persistent knowledge base (Qian et al., 1 Aug 2025). "Agent-as-a-Graph" formalizes agent and tool retrieval as a bipartite knowledge graph over parent agents and tool nodes, improving Recall@5 and nDCG@5 in large MCP ecosystems (Nizar et al., 22 Nov 2025). Enterprise deployment work such as IBM’s CUGA emphasizes planner–executor graphs, provenance, read-only modes, and domain-specific benchmarks when moving from prototype to business process automation (Shlomov et al., 27 Oct 2025).
This secondary usage does not redefine GuruAgents in the narrow investment sense. It suggests, however, that the term has become a convenient shorthand for prompt-guided agents whose specialization is expressed through structured prompts, tool registries, deterministic or semi-deterministic workflows, and explicit operational controls.
6. Limitations, misconceptions, and future directions
A common misconception is that GuruAgents are equivalent to a general multi-agent framework. In the original paper, they are five single-persona investment agents sharing one backbone model and a family of closely related deterministic pipelines, not a general orchestration substrate (Kim et al., 2 Oct 2025). A second misconception is that their results demonstrate universal superiority of LLM-guided investing. The reported evidence is a proof-of-concept on a short horizon, a single large-cap universe, and a specific market regime (Kim et al., 2 Oct 2025).
The original paper identifies several limits. The backtest spans less than two years, uses only NASDAQ-100 constituents, and does not deeply discuss survivorship bias or look-ahead handling. The gurus’ philosophies were historically developed for broader universes, often including smaller-cap and non-technology firms, so direct transfer to a mega-cap growth-heavy index may distort doctrinal fidelity. The study also does not include explicit ablations that remove tools, prompt sections, or scoring components one by one (Kim et al., 2 Oct 2025).
Future directions proposed in the paper include extending the approach to other asset classes, global and smaller-cap universes, formal evaluation of “philosophical fidelity,” finer-grained prompt ablations, and ensembles or negotiation protocols among multiple GuruAgents (Kim et al., 2 Oct 2025). A plausible implication is that the research program now splits in two directions: one line investigates GuruAgents as systematic-investing personas, while another treats “GuruAgents-like” systems as a broader engineering pattern for specialized LLM agents across web, enterprise, and software environments (Ma et al., 16 Feb 2026).