arxiv: 2603.13942 · v3 · submitted 2026-03-14 · 💱 q-fin.GN

Recognition: no theorem link

AI Agents in Financial Markets: Architecture, Applications, and Systemic Implications

Hui Gong

Authors on Pith no claims yet

Pith reviewed 2026-05-15 12:20 UTC · model grok-4.3

classification 💱 q-fin.GN

keywords AI agentsfinancial marketssystemic riskagent-based modelingfinancial automationmarket architectureAI governancebounded autonomy

0 comments

The pith

The systemic implications of AI in finance depend more on how agent architectures are distributed, coupled, and governed across institutions than on the intelligence of individual models.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper builds an integrative framework for agentic finance, in which autonomous AI systems perform perception, reasoning, strategy generation, and execution in markets. It introduces a stylized Agentic Financial Market Model that connects five agent design parameters to outcomes such as efficiency, liquidity resilience, volatility, and systemic risk. The core claim is that governance choices about how these agents interact matter more for stability than raw predictive power. An exploratory event study of capability disclosures illustrates one expectations channel. The near-term outlook the paper identifies is bounded autonomy, with agents operating as supervised modules inside human processes.

Core claim

The paper proposes a four-layer architecture for financial AI agents and the Agentic Financial Market Model (AFMM), a stylized agent-based representation that links parameters such as autonomy depth, heterogeneity, execution coupling, infrastructure concentration, and supervisory observability to market-level outcomes including efficiency, liquidity resilience, volatility, and systemic risk. It concludes that systemic implications depend less on model intelligence alone than on how agent architectures are distributed, coupled, and governed across institutions, with bounded autonomy as the most plausible near-term equilibrium.

What carries the argument

The Agentic Financial Market Model (AFMM), a stylized agent-based representation that maps agent design parameters to market outcomes.

If this is right

Bounded autonomy, in which AI agents function as supervised co-pilots and constrained execution modules, becomes the dominant near-term equilibrium.
Market efficiency and liquidity resilience scale with the heterogeneity and coupling patterns of deployed agents.
Systemic risk rises when infrastructure concentration increases without corresponding gains in supervisory observability.
Heterogeneous market repricing follows public disclosures of AI-agent capabilities, consistent with an expectations channel.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Regulators could shift emphasis from approving individual models to setting standards for agent interoperability and observability interfaces.
High infrastructure concentration may create new single points of failure that traditional stress tests do not capture.
Empirical work could test the model by examining order-flow data around periods when institutions announce changes in agent coupling.

Load-bearing premise

The five stylized parameters of the Agentic Financial Market Model can be linked to observable market outcomes without requiring full empirical validation of the model itself.

What would settle it

Observe whether markets that adopt high execution coupling and low supervisory observability across many institutions experience materially higher volatility or liquidity fragility than markets that retain lower coupling and higher observability.

Figures

Figures reproduced from arXiv: 2603.13942 by Hui Gong.

**Figure 2.** Figure 2: A four-layer architecture for AI agents in financial markets. [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗

**Figure 3.** Figure 3: Conceptual structure of the Agentic Financial Market Model (AFMM). [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗

read the original abstract

Recent advances in large language models, tool-using agents, and financial machine learning are shifting financial automation from isolated prediction tasks to integrated decision systems that can perceive information, reason over objectives, and generate or execute actions. This paper develops an integrative framework for analysing agentic finance: financial market environments in which autonomous or semi-autonomous AI systems participate in information processing, decision support, monitoring, and execution workflows. The analysis proceeds in three steps. First, the paper proposes a four-layer architecture of financial AI agents covering data perception, reasoning engines, strategy generation, and execution with control. Second, it introduces the Agentic Financial Market Model (AFMM), a stylised agent-based representation linking agent design parameters such as autonomy depth, heterogeneity, execution coupling, infrastructure concentration, and supervisory observability to market-level outcomes including efficiency, liquidity resilience, volatility, and systemic risk. Third, it presents an illustrative empirical application based on event studies of AI-agent capability disclosures and heterogeneous market repricing. The paper argues that the systemic implications of AI in finance depend less on model intelligence alone than on how agent architectures are distributed, coupled, and governed across institutions. The empirical application is intentionally exploratory: it does not validate the full AFMM, but shows how one observable expectations channel can be studied using public data. In the near term, the most plausible equilibrium is bounded autonomy, in which AI agents operate as supervised co-pilots, monitoring systems, and constrained execution modules embedded within human decision processes.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper gives a clean four-layer architecture for financial AI agents and introduces AFMM as a way to link design choices to market outcomes, but the mappings stay at the level of stylized parameters without explicit derivations or simulations.

read the letter

The main takeaway is that this paper organizes recent AI agent work into a four-layer structure (data perception, reasoning, strategy, execution) and proposes the Agentic Financial Market Model to connect parameters like autonomy depth, heterogeneity, execution coupling, infrastructure concentration, and supervisory observability to outcomes such as efficiency and systemic risk. The central argument is that distribution and governance of these agents matter more than raw model power alone, with bounded autonomy as the likely near-term setup. That framing is new as an integrative proposal for this domain. The exploratory event study on AI capability disclosures and market repricing uses public data in a straightforward way and shows one expectations channel without overclaiming validation. The synthesis of LLM agents with market microstructure concerns is done cleanly and stays practical. The soft spots sit in the AFMM linkages. No equations, functional forms, or simulation protocol appear for how the five parameters actually drive the market-level results, so the connections remain narrative rather than derived or tested. The empirical section is labeled exploratory and does not validate the full model, which means the strongest claims rest on the conceptual setup. This is not a fatal issue for a framework paper, but it does leave the parameter-to-outcome story open to the charge that parameters are fitted to stylized facts. The work is aimed at readers in fintech regulation, institutional design, or agent-based modeling who want a map of the space rather than a finished empirical test. It is coherent on its own terms and shows honest engagement with the literature, so it deserves a serious referee who can ask for more explicit dynamics or sensitivity checks. I would send it to peer review.

Referee Report

2 major / 2 minor

Summary. The paper develops an integrative framework for agentic finance in which autonomous or semi-autonomous AI systems participate in information processing, decision support, monitoring, and execution. It proposes a four-layer architecture (data perception, reasoning engines, strategy generation, execution with control), introduces the stylised Agentic Financial Market Model (AFMM) that links five agent-design parameters (autonomy depth, heterogeneity, execution coupling, infrastructure concentration, supervisory observability) to market-level outcomes (efficiency, liquidity resilience, volatility, systemic risk), and presents an exploratory empirical illustration based on event studies of AI-agent capability disclosures. The central argument is that systemic implications depend less on model intelligence alone than on how architectures are distributed, coupled, and governed.

Significance. If the AFMM parameter-to-outcome mappings can be made explicit and tested, the framework would provide a useful conceptual bridge between agent-level design choices and market-level stability metrics, shifting analytical attention from raw predictive performance to governance and coupling structures. The exploratory empirical component usefully flags an observable expectations channel but does not itself establish the model's broader validity.

major comments (2)

[AFMM introduction] AFMM section: the five parameters are asserted to determine market outcomes more strongly than raw model intelligence, yet no equations, functional forms, simulation protocol, or sensitivity analysis are supplied to show how changes in autonomy depth, heterogeneity, execution coupling, etc., map onto efficiency, volatility, or systemic risk; the linkage therefore remains narrative rather than demonstrated.
[Empirical application] Empirical application section: the event-study illustration is explicitly described as non-validating of the full AFMM and relies on public data for one expectations channel only; this leaves the central claim about architecture-driven systemic implications without direct empirical support within the manuscript.

minor comments (2)

[Architecture section] The four-layer architecture description would be clearer with an accompanying diagram showing information flows and control loops between layers.
[AFMM section] Notation for the AFMM parameters is introduced but not consistently referenced in the discussion of market outcomes; a summary table linking each parameter to the claimed directional effects would improve readability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback and the recommendation for major revision. We agree that the AFMM would benefit from greater explicitness in its parameter linkages and will revise the manuscript accordingly while preserving its conceptual focus. We address each major comment below.

read point-by-point responses

Referee: [AFMM introduction] AFMM section: the five parameters are asserted to determine market outcomes more strongly than raw model intelligence, yet no equations, functional forms, simulation protocol, or sensitivity analysis are supplied to show how changes in autonomy depth, heterogeneity, execution coupling, etc., map onto efficiency, volatility, or systemic risk; the linkage therefore remains narrative rather than demonstrated.

Authors: The AFMM is presented as a stylised conceptual framework rather than a fully specified quantitative model, drawing on established principles from agent-based computational finance to highlight how architectural parameters can dominate raw intelligence in shaping market outcomes. We acknowledge that the current description is primarily qualitative. In the revised version we will add an explicit table mapping each of the five parameters to directional effects on efficiency, liquidity, volatility and systemic risk, together with a brief outline of a potential simulation protocol that future work could implement. This strengthens the linkage without converting the paper into a simulation study. revision: partial
Referee: [Empirical application] Empirical application section: the event-study illustration is explicitly described as non-validating of the full AFMM and relies on public data for one expectations channel only; this leaves the central claim about architecture-driven systemic implications without direct empirical support within the manuscript.

Authors: The manuscript already states that the event study is exploratory, limited to one observable expectations channel, and does not constitute validation of the full AFMM. The central claims about architecture-driven systemic implications rest on the theoretical AFMM framework. We will revise the empirical section and conclusion to further underscore these scope limitations and to indicate how subsequent empirical designs could test the broader parameter mappings. This maintains the paper's transparency about the current state of evidence. revision: partial

Circularity Check

1 steps flagged

AFMM defined as linking device makes parameter-to-outcome claim true by construction

specific steps

self definitional [Abstract]
"it introduces the Agentic Financial Market Model (AFMM), a stylised agent-based representation linking agent design parameters such as autonomy depth, heterogeneity, execution coupling, infrastructure concentration, and supervisory observability to market-level outcomes including efficiency, liquidity resilience, volatility, and systemic risk. ... The empirical application is intentionally exploratory: it does not validate the full AFMM"

The AFMM is defined as the representation that performs the linking between the listed parameters and outcomes. The strongest claim (systemic implications depend on architectures via these parameters) is therefore true by the model's definitional role rather than by any shown derivation, dynamics, or validated mapping; the exploratory empirical section explicitly disclaims full validation.

full rationale

The paper's central claim—that systemic implications depend on agent architecture distribution rather than model intelligence—rests on the AFMM as the linking model. However, the AFMM is introduced explicitly as a stylized representation whose purpose is to connect the five design parameters to the listed market outcomes. No equations, functional forms, simulation protocol, or independent derivation of the mapping are provided; the empirical section is stated to be exploratory and non-validating. This reduces the load-bearing linkage to a definitional assertion rather than a demonstrated result, producing partial circularity while leaving the architecture description and exploratory event study as independent content.

Axiom & Free-Parameter Ledger

5 free parameters · 2 axioms · 2 invented entities

The framework introduces design parameters and a new stylized model without independent empirical grounding beyond an exploratory study; standard agent-based modeling assumptions are invoked implicitly.

free parameters (5)

autonomy depth
Agent design parameter treated as input to market outcomes in AFMM
heterogeneity
Agent design parameter treated as input to market outcomes in AFMM
execution coupling
Agent design parameter treated as input to market outcomes in AFMM
infrastructure concentration
Agent design parameter treated as input to market outcomes in AFMM
supervisory observability
Agent design parameter treated as input to market outcomes in AFMM

axioms (2)

domain assumption Agent-based models can represent financial market dynamics from micro-level agent rules
Invoked to justify linking agent parameters to macro outcomes in AFMM
domain assumption Event-study methods on capability disclosures can isolate expectations channels
Basis for the illustrative empirical application

invented entities (2)

Agentic Financial Market Model (AFMM) no independent evidence
purpose: Stylized representation connecting agent design parameters to market-level outcomes
New model introduced to organize the analysis
Four-layer architecture (data perception, reasoning engines, strategy generation, execution with control) no independent evidence
purpose: Organizing framework for financial AI agents
New integrative structure proposed in the paper

pith-pipeline@v0.9.0 · 5560 in / 1673 out tokens · 47640 ms · 2026-05-15T12:20:21.234955+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

27 extracted references · 27 canonical work pages · 2 internal anchors

[1]

URLhttps://link.springer.com/article/10.1007/s104 62-025-11422-4

doi: 10.1007/s10462-025-11422-4. URLhttps://link.springer.com/article/10.1007/s104 62-025-11422-4. Inaki Aldasoro, Leonardo Gambacorta, Anton Korinek, Vatsala Shreeti, and Merlin Stein. Intel- ligent financial system: How AI is transforming finance. BIS Working Papers 1194, Bank for International Settlements,

work page doi:10.1007/s10462-025-11422-4
[2]

Carlos Betancourt and Wen-Hui Chen

doi: 10.1080/14697680701381228. Carlos Betancourt and Wen-Hui Chen. Deep reinforcement learning for portfolio management of markets with a dynamic number of assets.Expert Systems with Applications, 164:114002,

work page doi:10.1080/14697680701381228
[3]

URLhttps://www.sciencedirect.com/science/article/pi i/S0957417420307776

doi: 10.1016/j.eswa.2020.114002. URLhttps://www.sciencedirect.com/science/article/pi i/S0957417420307776. Juan Carlos Crisanto, Cris Benson Leuterio, Jermy Prenio, and Jeffery Yong. Regulating AI in the financial sector: Recent developments and main challenges. FSI Insights on Policy Implementa- tion 63, Bank for International Settlements,

work page doi:10.1016/j.eswa.2020.114002 2020
[4]

32 European Union

URLhttps://www.bis.org/fsi/publ/insig hts63.htm. 32 European Union. Regulation (EU) 2022/2554 of the european parliament and of the council of 14 december 2022 on digital operational resilience for the financial sector (DORA),

work page 2022
[5]

European Union

URL https://eur-lex.europa.eu/eli/reg/2022/2554/oj. European Union. Regulation (EU) 2023/1114 of the european parliament and of the council of 31 may 2023 on markets in crypto-assets (MiCA),

work page 2022
[6]

European Union

URLhttps://eur-lex.europa.eu/eli/r eg/2023/1114/oj. European Union. Regulation (EU) 2024/1689 of the european parliament and of the council of 13 june 2024 laying down harmonised rules on artificial intelligence (AI Act),

work page 2023
[7]

URLhttps: //eur-lex.europa.eu/eli/reg/2024/1689/oj. J. Doyne Farmer and Duncan Foley. The economy needs agent-based modelling.Nature, 460(7256): 685–686,

work page 2024
[8]

URLhttps://www.nature.com/articles/460685a

doi: 10.1038/460685a. URLhttps://www.nature.com/articles/460685a. Financial Stability Board. The financial stability implications of artificial intelligence. Technical report, Financial Stability Board,

work page doi:10.1038/460685a
[10]

Joel Hasbrouck

URLhttps://arxiv.org/abs/2406.19966. Joel Hasbrouck. High-frequency quoting: Short-term volatility in bids and offers.Journal of Fi- nancial and Quantitative Analysis, 53(2):613–641,

work page arXiv
[11]

doi: 10.1017/S0022109017001053. URL https://www.cambridge.org/core/journals/journal-of-financial-and-quantitative-a nalysis/article/highfrequency-quoting-shortterm-volatility-in-bids-and-offers/B4 BD84DAD1B0C6577EA51A47C0A8B70C. J.B.Heaton, NicholasG.Polson, andJanHendrikWitte. Deeplearningforfinance: Deepportfolios. Applied Stochastic Models in Business ...

work page doi:10.1017/s0022109017001053
[12]

Steven L

doi: 10.1002/asmb.2209. URLhttps://econpapers.repec.org/RePEc:wly:apsmbi:v:33:y:2017:i:1:p:3-12. Terrence Hendershott, Charles M. Jones, and Albert J. Menkveld. Does algorithmic trading improve liquidity?The Journal of Finance, 66(1):1–33,

work page doi:10.1002/asmb.2209 2017
[13]

URLhttps://research.vu.nl/en/publications/does-algorithmic-trading-improve-liq uidity

doi: 10.1111/j.1540-6261.2010.01624.x. URLhttps://research.vu.nl/en/publications/does-algorithmic-trading-improve-liq uidity. Cars Hommes.Behavioral Rationality and Heterogeneous Expectations in Complex Economic Sys- tems. Cambridge University Press, Cambridge,

work page doi:10.1111/j.1540-6261.2010.01624.x 2010
[14]

URL https://www.cambridge.org/core/product/identifier/9781139094276/type/book

doi: 10.1017/CBO9781139094276. URL https://www.cambridge.org/core/product/identifier/9781139094276/type/book. Satyadhar Joshi. Agentic RIAs: Strengthening US financial stability through AI architecture, regulation, and systemic integration. Preprints.org preprint,

work page doi:10.1017/cbo9781139094276
[15]

Andrei A

URLhttps://www.preprint s.org/manuscript/202602.0975/v1. Andrei A. Kirilenko, Albert S. Kyle, Mehrdad Samadi, and Tugkan Tuzun. The flash crash: High- frequency trading in an electronic market.Journal of Finance, 72(3):967–998,

work page arXiv
[16]

URLhttps://www.sciencedirect.co m/science/article/pii/S1574002105020241

doi: 10.1016/S1574-0021(05)02024-1. URLhttps://www.sciencedirect.co m/science/article/pii/S1574002105020241. Zhaowei Liu, Xin Guo, Hongrui Cai, Xian Li, and Mingyang Jiang. Fin-R1: A large language model for financial reasoning through reinforcement learning.arXiv preprint arXiv:2503.16252,

work page doi:10.1016/s1574-0021(05)02024-1
[17]

Marcos Lopez de Prado.Machine Learning for Asset Managers

URLhttps://arxiv.org/abs/2503.16252. Marcos Lopez de Prado.Machine Learning for Asset Managers. Cambridge University Press, Cambridge,

work page arXiv
[18]

Maxim Massenkoff and Peter McCrory

doi: 10.1017/9781108883658. Maxim Massenkoff and Peter McCrory. Labor market impacts of AI: A new measure and early evidence. Technical report, Anthropic, March

work page doi:10.1017/9781108883658
[19]

URLhttps://pubmed.ncbi

doi: 10.1109/72.935097. URLhttps://pubmed.ncbi. nlm.nih.gov/18249919/. OECD. Generative artificial intelligence in finance. OECD Artificial Intelligence Papers 9, OECD Publishing,

work page doi:10.1109/72.935097
[20]

O’Brien, Carrie J

doi: 10.1145/3586183.3606763. Long Qian, Xiaohao Peng, Yuchen Wang, Eric Alcaide, Troy Fang, Tanmoy Chakraborty, Xuhui Zhou, and Chaomei Chen. When agents trade: Live multi-market trading benchmark for LLM agents.arXiv preprint arXiv:2510.11695,

work page doi:10.1145/3586183.3606763
[21]

URLhttps://arxiv.org/abs/2510.11695. Reuters. Ibm posts steepest daily drop since 2000 after anthropic says AI can modernize COBOL, February

work page arXiv 2000
[22]

Published February 24,

URLhttps://www.reuters.com/business/ibm-posts-steepest-daily-dro p-since-2000-after-anthropic-says-ai-can-modernize-2026-02-24/. Published February 24,

work page 2000
[23]

URLhttps://www.sciencedirect.com/scie nce/article/pii/S1546221825010938

doi: 10.32604/cmc.2025.069678. URLhttps://www.sciencedirect.com/scie nce/article/pii/S1546221825010938. Shijie Wu, Ozan Irsoy, Steven Lu, Vadim Dabravolski, Mark Dredze, Sebastian Gehrmann, Prab- hanjan Kambadur, David Rosenberg, and Gideon Mann. BloombergGPT: A large language model for finance.arXiv preprint arXiv:2303.17564,

work page doi:10.32604/cmc.2025.069678 2025
[24]

BloombergGPT: A Large Language Model for Finance

URLhttps://arxiv.org/abs/2303.17564. 34 Yingqian Wu, Qiushi Wang, Zefei Long, Rong Ye, Zhongtian Lu, Xianyin Zhang, Bingxuan Li, Wei Chen, Liwen Zhang, and Zhongyu Wei. FinTeam: A multi-agent collaborative intelligence systemforcomprehensivefinancialscenarios.arXiv preprint arXiv:2507.10448,

work page internal anchor Pith review Pith/arXiv arXiv
[25]

Zhaohan Xi, Wei Chen, Xiaoxu Guo, Weizhe He, and Yifei Ding

URLhttps: //arxiv.org/abs/2507.10448. Zhaohan Xi, Wei Chen, Xiaoxu Guo, Weizhe He, and Yifei Ding. The rise and potential of large language model based agents: A survey.arXiv preprint arXiv:2309.07864,

work page arXiv
[26]

The Rise and Potential of Large Language Model Based Agents: A Survey

URLhttps: //arxiv.org/abs/2309.07864. Yijia Xiao, Edward Sun, Di Luo, and Wei Wang. TradingAgents: Multi-agent LLM financial trading framework. InThe First Multi-Agent AI in the Real World Workshop at AAAI 2025,

work page internal anchor Pith review Pith/arXiv arXiv 2025
[27]

Hongyang Yang, Xiao-Yang Liu, and Christina Wang

URLhttps://proceedings.neurips.cc /paper_files/paper/2024/hash/adb1d9fa8be4576d28703b396b82ba1b-Abstract-Datasets_ and_Benchmarks_Track.html. Hongyang Yang, Xiao-Yang Liu, and Christina Wang. FinGPT: Open-source financial large lan- guage models.arXiv preprint arXiv:2306.06031,

work page arXiv 2024
[28]

Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao

URL https://neurips.cc/virtual/2025/poster/116676. Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. ReAct: Synergizing reasoning and acting in language models. InThe Eleventh International Conference on Learning Representations,

work page 2025