AI Agents in Financial Markets: Architecture, Applications, and Systemic Implications金融市场中的人工智能代理：架构、应用与系统性影响

The relevant literature is broader than the recent wave of financial LLM papers. For the purposes of this paper, it is useful to organise prior work into six streams: (i) agent-based finance and heterogeneous-expectations models, (ii) market microstructure and automated trading, (iii) financial LLMs and reasoning benchmarks, (iv) agent architectures and multi-agent systems, (v) AI in trading and investment management, and (vi) regulation and systemic-risk analysis.相关文献的范畴比近期涌现的金融大语言模型论文更为广泛。就本文而言，将既往研究归纳为六个流派是有益的：(i) 基于代理的金融与异质预期模型，(ii) 市场微观结构与自动交易，(iii) 金融大语言模型与推理基准，(iv) 代理架构与多代理系统，(v) 交易与投资管理中的人工智能，以及 (vi) 监管与系统性风险分析。

Agent-based finance and heterogeneous expectations. A first intellectual antecedent comes from agent-based finance and the broader literature on heterogeneous expectations. Classic surveys and syntheses show how markets populated by heterogeneous agents can generate endogenous volatility, bubbles, crashes, and persistent deviations from benchmark rational-expectations dynamics (LeBaron, 2006; Farmer and Foley, 2009; Hommes, 2013). This literature is directly relevant to the present paper because the AFMM is ultimately concerned with how diversity, coupling, and interaction among heterogeneous decision systems shape market outcomes. Recent LLM-based market simulations such as ASFM and TwinMarket can be interpreted as a new technological branch of this older heterogeneous-agent tradition, now populated by AI-enabled rather than purely stylised bounded-rational agents (Gao et al., 2024; Yang et al., 2025).基于代理的金融与异质预期。第一个智力渊源来自基于代理的金融及更广泛的异质预期文献。经典的综述与综合研究表明，由异质代理构成的市场如何产生内生波动、泡沫、崩盘以及对基准理性预期动态的持续偏离（LeBaron, 2006; Farmer and Foley, 2009; Hommes, 2013）。该文献与本文直接相关，因为 AFMM 最终关注的是异质决策系统之间的多样性、耦合和交互如何塑造市场结果。近期基于大语言模型的市场模拟（如 ASFM 和 TwinMarket）可以被解读为这一古老异质代理传统的新技术分支，现在由人工智能驱动的代理取代了纯粹风格化的有限理性代理（Gao et al., 2024; Yang et al., 2025）。

Market microstructure and automated trading. A second stream comes from market microstructure and the literature on automated and algorithmic trading. This work establishes that automated agents have already altered liquidity provision, price discovery, and intraday volatility long before the arrival of contemporary LLM agents (Avellaneda and Stoikov, 2008; Hendershott et al., 2011; Hasbrouck, 2018; Kirilenko et al., 2017). The relevance of this literature is twofold. First, it provides evidence that automation can materially change market quality and market dynamics. Second, it supplies a benchmark against which agentic finance can be compared: the move from algorithmic trading to agentic finance is not the first interaction between automation and markets, but it is a deeper extension from automated execution toward automated workflow coordination.市场微观结构与自动交易。第二个流派来自市场微观结构以及自动和算法交易文献。这项工作确立了这样一个事实：早在当代大语言模型代理出现之前，自动代理就已经改变了流动性供给、价格发现和日内波动（Avellaneda and Stoikov, 2008; Hendershott et al., 2011; Hasbrouck, 2018; Kirilenko et al., 2017）。该文献的相关性体现在两个方面：首先，它提供了自动化可以实质性改变市场质量和市场动态的证据；其次，它提供了一个可以衡量代理金融的基准：从算法交易向代理金融的演进并非自动化与市场的首次交互，而是从自动执行向自动工作流协调的更深层扩展。

Financial LLMs and reasoning benchmarks. A third stream examines domain adaptation and evaluation of large language models in financial contexts. Early work such as BloombergGPT demonstrated that large-scale domain-specific pretraining can significantly improve performance on financial NLP tasks including sentiment classification, information extraction, and document understanding (Wu et al., 2023). Subsequent work such as FinGPT expanded this approach by proposing open-source pipelines for financial language models and demonstrating their use in financial analysis and forecasting-related workflows (Yang et al., 2023). Benchmark studies such as FinBen further highlight the uneven performance of LLMs across financial reasoning tasks, emphasizing that financial competence cannot be reduced to generic language understanding alone (Xie et al., 2024). More recent models such as Fin-R1 explore reinforcement-learning-based improvements in structured financial reasoning, particularly for tasks such as compliance analysis and financial question answering (Liu et al., 2025).金融大语言模型与推理基准。第三个流派考察了金融背景下大语言模型的领域适配与评估。早期的工作（如 BloombergGPT）证明，大规模特定领域预训练可以显著提升金融自然语言处理任务的性能，包括情感分类、信息提取和文档理解（Wu et al., 2023）。随后的工作（如 FinGPT）通过提出金融语言模型的开源流水线并展示其在金融分析和预测相关工作流中的应用，扩展了这一方法（Yang et al., 2023）。FinBen 等基准研究进一步强调了大语言模型在金融推理任务中表现不均衡的情况，强调金融能力不能仅简化为通用语言理解（Xie et al., 2024）。更近期的模型（如 Fin-R1）探索了基于强化学习的结构化金融推理改进，特别是在合规分析和金融问答等任务中（Liu et al., 2025）。

Agent architectures and multi-agent systems. A fourth stream focuses on the architectural foundations of AI agents. General frameworks such as ReAct demonstrate how reasoning and acting can be integrated in language-model workflows (Yao et al., 2023), while generative-agent frameworks illustrate how memory, planning, and environment interaction can be combined in persistent agent-based environments (Park et al., 2023). Survey work further systematises this design space by analysing memory structures, tool usage, planning mechanisms, and multi-agent coordination (Xi et al., 2023). Building on these foundations, finance-specific work proposes collaborative multi-agent systems for financial analysis and investment reasoning. FinTeam, for example, models financial analysis as a coordinated interaction between specialised analyst agents responsible for macro interpretation, risk evaluation, and portfolio reasoning (Wu et al., 2025).代理架构与多代理系统。第四个流派聚焦于人工智能代理的架构基础。通用框架（如 ReAct）展示了推理与行动如何整合在语言模型工作流中（Yao et al., 2023），而生成式代理框架则阐明了记忆、规划和环境交互如何结合在持久的基于代理的环境中（Park et al., 2023）。综述性工作通过分析记忆结构、工具使用、规划机制和多代理协调，进一步系统化了这一设计空间（Xi et al., 2023）。在此基础上，金融专项研究提出了用于金融分析和投资推理的协作式多代理系统。例如，FinTeam 将金融分析建模为负责宏观解读、风险评估和投资组合推理的专业分析师代理之间的协调交互（Wu et al., 2025）。

AI in trading and investment management. A fifth stream studies AI deployment in trading, allocation, and investment support. Earlier work in this area includes reinforcement-learning-based trading and portfolio control, as well as early deep-learning approaches to portfolio design and financial prediction (Moody and Saffell, 2001; Heaton et al., 2017; Betancourt and Chen, 2021; Lopez de Prado, 2020). More recent research moves toward LLM-based trading agents and multi-agent trading environments. TradingAgents proposes a collaborative architecture composed of analyst, trader, and risk-management agents (Xiao et al., 2025), while When Agents Trade introduces a live multi-market benchmark for evaluating LLM trading agents across equities and crypto markets (Qian et al., 2025). Together, these studies show how AI applications are shifting from narrow signal generation toward integrated decision pipelines that include interpretation, strategy construction, and execution support.交易与投资管理中的人工智能。第五个流派研究人工智能在交易、资产配置和投资支持中的部署。该领域的早期工作包括基于强化学习的交易和投资组合控制，以及用于投资组合设计和金融预测的早期深度学习方法（Moody and Saffell, 2001; Heaton et al., 2017; Betancourt and Chen, 2021; Lopez de Prado, 2020）。更近期的研究转向基于大语言模型的交易代理和多代理交易环境。TradingAgents 提出了一种由分析师、交易员和风险管理代理组成的协作架构（Xiao et al., 2025），而 When Agents Trade 则引入了一个用于评估股票和加密市场大语言模型交易代理的实时多市场基准（Qian et al., 2025）。这些研究共同展示了人工智能应用如何从狭义的信号生成转向包含解读、策略构建和执行支持的集成决策流水线。

Regulation and systemic risk. The final stream focuses on governance, regulation, and systemic implications. Reports from the BIS, OECD, and the Financial Stability Board highlight several emerging concerns, including model opacity, concentration of third-party providers, operational resilience, and the potential for correlated machine-mediated behaviour across financial institutions (OECD, 2023; OECD, 2024; Financial Stability Board, 2024; Aquilina et al., 2025). This literature is crucial because it widens the unit of analysis from firm-level productivity to financial stability, supervisory capacity, and market design.监管与系统性风险。最后一个流派聚焦于治理、监管和系统性影响。国际清算银行、经合组织和金融稳定理事会的报告强调了几个新兴担忧，包括模型不透明、第三方提供商的集中化、运营韧性，以及金融机构间可能出现的机器中介行为的相关性（OECD, 2023; OECD, 2024; Financial Stability Board, 2024; Aquilina et al., 2025）。该文献至关重要，因为它将分析单元从企业层面的生产力扩展到了金融稳定性、监管能力和市场设计。

Taken together, these six research streams demonstrate that the technological and institutional foundations of AI-driven finance are rapidly evolving. However, three analytical gaps remain particularly important.综合来看，这六个研究流表明人工智能驱动的金融技术和制度基础正在快速演进。然而，仍有三个分析缺口尤为重要。

First, the literature lacks a finance-specific architectural framework for AI agents. Most existing work focuses either on general agent architectures or on specific financial applications. As a result, there is little consensus on how to decompose financial AI agents into modular components that reflect the structure of financial decision-making. Without such an architecture, it is difficult to systematically compare agent designs across institutions or applications.首先，文献缺乏针对人工智能代理的金融专用架构框架。大多数现有工作要么关注通用代理架构，要么关注特定的金融应用。因此，对于如何将金融人工智能代理分解为反映金融决策结构的模块化组件，目前尚无共识。没有这样的架构，就很难在机构或应用之间系统地比较代理设计。

Second, the connection between agent design and financial market outcomes remains underdeveloped. The agent-based finance and market-microstructure literatures already show that heterogeneous automated agents can shape liquidity, volatility, and price discovery. However, these insights are rarely connected to the architecture of contemporary AI systems. Many recent studies still evaluate AI at the level of isolated tasks such as classification, summarisation, or signal generation. Questions related to liquidity provision, volatility amplification, and behavioural correlation therefore remain insufficiently connected to the technical design of modern agentic systems.其次，代理设计与金融市场结果之间的联系尚不完善。基于代理的金融和市场微观结构文献已经表明，异质自动代理可以塑造流动性、波动性和价格发现。然而，这些见解很少与当代人工智能系统的架构联系起来。许多近期研究仍仅在分类、摘要或信号生成等孤立任务层面评估人工智能。因此，与流动性供给、波动性放大和行为相关性相关的问题，仍未充分关联到现代代理系统的技术设计上。

Third, discussions of systemic risk are often disconnected from the micro-structure of AI systems. Policy-oriented research frequently identifies risks such as model homogeneity, third-party dependencies, and correlated decision-making. Yet these discussions are rarely grounded in a clear model of how agentic systems are constructed, how they process information, and how their outputs translate into financial actions. This limits the ability of both researchers and regulators to link system-level concerns to specific design choices.第三，关于系统性风险的讨论往往与人工智能系统的微观结构脱节。政策导向的研究经常识别出模型同质性、第三方依赖和相关决策等风险。然而，这些讨论很少建立在清晰的模型之上，即代理系统是如何构建的、它们如何处理信息，以及它们的输出如何转化为金融行动。这限制了研究人员和监管者将系统层面的担忧与具体设计选择联系起来的能力。

Against this background, the objective of this paper is not simply to review emerging research, but to develop an integrated theoretical framework for what we term agentic finance. The central contribution is to bridge the gap between the technical design of AI agents and the economic structure of financial markets.在此背景下，本文的目标不仅仅是回顾新兴研究，而是为我们所称的“代理金融”开发一个综合理论框架。其核心贡献在于架起人工智能代理技术设计与金融市场经济结构之间的桥梁。

More specifically, the paper makes three theoretical contributions:更具体地说，本文做出了三项理论贡献：

1. (i)

   it proposes a modular four-layer architecture of AI financial agents that decomposes agent systems into data perception, reasoning, strategy generation, and execution with control;提出了一个模块化的金融人工智能代理四层架构，将代理系统分解为数据感知、推理、策略生成以及受控执行；

2. (ii)

   it develops an Agentic Financial Market Model (AFMM) that links micro-level agent design choices to macro-level market outcomes such as efficiency, liquidity provision, volatility dynamics, herding behaviour, and market concentration;开发了一个“代理金融市场模型”（AFMM），将微观层面的代理设计选择与宏观层面的市场结果（如效率、流动性供给、波动性动态、羊群行为和市场集中度）联系起来；

3. (iii)

   it formulates a set of mechanism-based propositions that connect technical features of agentic systems—such as model heterogeneity, execution coupling, and infrastructure concentration—to potential financial-system consequences.制定了一系列基于机制的命题，将代理系统的技术特征（如模型异构性、执行耦合度和基础设施集中度）与潜在的金融系统后果联系起来。

The key theoretical move of the paper is therefore to treat AI agents not merely as decision-support tools but as market actors embedded in shared infrastructures. Once framed in this way, the relevant analytical variables extend beyond model accuracy or benchmark performance. They include heterogeneity of models and data sources, coupling between decision and execution layers, concentration of technical infrastructure, the degree of delegated autonomy, and the visibility of agent behaviour to supervisors.因此，本文的关键理论举措是将人工智能代理不仅视为决策支持工具，而且视为嵌入共享基础设施中的市场参与者。一旦以这种方式构建，相关的分析变量就超出了模型准确性或基准性能的范畴。它们包括模型和数据源的异构性、决策层与执行层之间的耦合、技术基础设施的集中度、授权自主的程度，以及监管者对代理行为的可视度。

By explicitly linking these variables to financial market outcomes, the framework developed in this paper aims to provide a foundation for future empirical research on the economic and systemic implications of agentic financial systems.通过将这些变量与金融市场结果明确挂钩，本文开发的框架旨在为未来关于代理金融系统经济和系统性影响的实证研究奠定基础。

The next section places this framework in historical context. Doing so clarifies why the move from algorithmic and machine-learning finance to agentic finance is not simply an incremental improvement in model quality, but a shift in the scope of automation itself.下一节将把该框架置于历史背景中。这样做可以阐明为什么从算法和机器学习金融向代理金融的转变，不仅仅是模型质量的渐进式改进，而是自动化范畴本身的转变。

The first generation of financial AI was centred on algorithmic finance. Its defining feature was the automation of rule-based execution and market interaction. In this paradigm, systems were designed to follow explicit instructions under pre-specified conditions: for example, splitting orders, optimising execution schedules, or responding to microstructure signals in electronic markets (Avellaneda and Stoikov, 2008; Hendershott et al., 2011; Hasbrouck, 2018; Kirilenko et al., 2017). The economic logic of this generation was primarily operational. Automation improved speed, reduced execution costs, and enabled more systematic participation in increasingly electronic markets.金融人工智能的第一世代以算法金融为中心。其定义特征是规则驱动的执行和市场交互的自动化。在此范式下，系统被设计为在预先指定的条件下遵循明确指令：例如，拆分订单、优化执行计划或响应电子市场中的微观结构信号（Avellaneda and Stoikov, 2008; Hendershott et al., 2011; Hasbrouck, 2018; Kirilenko et al., 2017）。这一世代的经济逻辑主要是操作性的。自动化提高了速度，降低了执行成本，并实现了在日益电子化的市场中更系统的参与。

Although algorithmic finance transformed market microstructure, its scope remained limited. These systems did not typically “understand” news, reason over changing objectives, or integrate heterogeneous information sources. Their intelligence lay in disciplined execution rather than adaptive interpretation. In other words, the automated object was the rule, not the broader decision workflow. This first generation therefore established the infrastructure of automated finance without yet creating genuinely flexible or context-sensitive decision systems.尽管算法金融改变了市场微观结构，但其范畴仍然有限。这些系统通常不会“理解”新闻、围绕不断变化的目标进行推理，也不会整合异构信息源。它们的智能在于纪律严明的执行，而非适应性的解读。换言之，自动化的对象是规则，而非更广泛的决策工作流。因此，第一世代建立了自动化金融的基础设施，但尚未创建真正灵活或对环境敏感的决策系统。

The second generation can be described as machine-learning finance. Here the main object of automation shifted from execution rules to predictive tasks. Financial institutions increasingly used statistical learning and machine-learning methods for return prediction, volatility forecasting, credit scoring, fraud detection, portfolio analytics, and risk monitoring (Moody and Saffell, 2001; Heaton et al., 2017; Betancourt and Chen, 2021; Lopez de Prado, 2020). Relative to the first generation, this represented a major expansion in analytical capability. Systems were no longer limited to explicit rules; they could infer patterns from data and generate probabilistic assessments that improved classification, ranking, forecasting, and portfolio construction.第二世代可以描述为机器学习金融。在这里，自动化的主要对象从执行规则转移到了预测任务。金融机构越来越多地将统计学习和机器学习方法用于收益预测、波动性预测、信用评分、欺诈检测、投资组合分析和风险监控（Moody and Saffell, 2001; Heaton et al., 2017; Betancourt and Chen, 2021; Lopez de Prado, 2020）。相对于第一世代，这代表了分析能力的重大扩展。系统不再局限于明确的规则；它们可以从数据中推断模式并生成概率评估，从而改进分类、排序、预测和投资组合构建。

Yet machine-learning finance was still largely model-centric. Most systems were designed to produce a signal, score, forecast, or recommendation at a specific point in the workflow. Human actors remained responsible for integrating that output into a broader sequence of tasks: gathering contextual information, interpreting significance, checking constraints, and deciding how or whether to act. Even when machine-learning models were highly accurate on narrow tasks, they did not themselves manage the end-to-end process through which financial actions are generated and implemented. Their contribution was therefore substantial, but still partial.然而，机器学习金融在很大程度上仍然是以模型为中心的。大多数系统被设计为在工作流的特定点产生信号、分数、预测或建议。人类参与者仍然负责将该输出整合到更广泛的任务序列中：收集背景信息、解读重要性、检查约束，并决定如何行动或是否行动。即使机器学习模型在狭窄任务上非常准确，它们本身也不会管理金融行动生成和实施的端到端过程。因此，它们的贡献是实质性的，但仍然是局部的。

The third generation, now emerging, is agentic finance. In this generation, AI systems extend beyond isolated prediction or execution tasks and begin to operate across a larger portion of the financial workflow. Advances in large language models, retrieval systems, planning modules, memory architectures, and tool-using agents have made it increasingly feasible to build systems that can absorb heterogeneous inputs, reason across intermediate steps, coordinate external tools, maintain state over time, and recommend or initiate actions in sequential environments (Yao et al., 2023; Park et al., 2023; Xi et al., 2023).正在兴起的第三世代是代理金融。在这一世代中，人工智能系统超越了孤立的预测或执行任务，开始在金融工作流的更大比例中运行。大语言模型、检索系统、规划模块、记忆架构和工具使用代理的进步，使得构建能够吸收异构输入、跨中间步骤进行推理、协调外部工具、随时间保持状态，并在序列环境中推荐或发起行动的系统变得越来越可行（Yao et al., 2023; Park et al., 2023; Xi et al., 2023）。

What distinguishes agentic finance from earlier generations is therefore not only improved intelligence, but workflow integration. An agent is not merely a forecasting model or an execution algorithm. It is a goal-directed system that can transform inputs into market-relevant decision objects under explicit objectives and constraints. In financial contexts, those decision objects may include trade ideas, portfolio reallocations, anomaly alerts, hedging proposals, compliance flags, or monitoring outputs. The critical point is that the system participates in the chain that links information to action.因此，将代理金融与早期世代区分开来的不仅仅是智能的提升，还有工作流的整合。代理不仅仅是一个预测模型或执行算法，它是一个目标导向的系统，能够在明确的目标和约束下将输入转化为与市场相关的决策对象。在金融背景下，这些决策对象可能包括交易想法、投资组合再平衡、异常警报、对冲建议、合规标记或监控输出。关键点在于，该系统参与了将信息与行动联系起来的链条。

This broader scope also changes the risk profile of AI in finance. In algorithmic finance, the main concerns centred on execution speed, market impact, and microstructure instability. In machine-learning finance, concerns expanded to include model risk, overfitting, and interpretability. In agentic finance, these issues remain relevant, but they are joined by additional concerns related to autonomy, coordination, infrastructure dependence, control loss, and the possibility of correlated system behaviour across institutions. As a result, the analytical challenge is no longer simply to ask whether AI improves local task performance. It is to ask how different forms of agentic design may reshape financial workflows and, ultimately, market dynamics.这种更广泛的范畴也改变了金融领域人工智能的风险状况。在算法金融中，主要担忧集中在执行速度、市场影响和微观结构不稳定上；在机器学习金融中，担忧扩展到包括模型风险、过拟合和可解释性；在代理金融中，这些问题依然相关，但又增加了与自主权、协调、基础设施依赖、控制丧失以及各机构间可能出现的相关系统行为相关的额外担忧。因此，分析挑战不再仅仅是询问人工智能是否改善了局部任务性能，而是询问不同形式的代理设计如何重塑金融工作流，并最终改变市场动态。

The transition across the three generations can be summarised as a shift from rule automation, to prediction automation, to workflow automation. This framing is important because it clarifies why agentic finance deserves separate analytical treatment. If AI systems increasingly operate across multiple stages of financial decision-making, then the relevant unit of analysis is no longer the isolated model, but the agent system as a structured combination of data perception, reasoning, strategy formation, and execution.跨越三个世代的转变可以总结为从规则自动化，到预测自动化，再到工作流自动化的演进。这一框架之所以重要，是因为它阐明了为什么代理金融值得进行单独的分析处理。如果人工智能系统日益在金融决策的多个阶段运行，那么相关的分析单元就不再是孤立的模型，而是作为数据感知、推理、策略形成和执行的结构化组合的代理系统。

Figure 1 illustrates this progression. The figure is intentionally stylised: in practice, the three generations overlap, and institutions often combine them within the same workflow. A modern financial system may still rely on algorithmic execution, machine-learning forecasts, and agentic reasoning simultaneously. Nevertheless, the generational framing is analytically useful because it highlights a real transformation in the locus of automation. The central question is no longer whether AI can assist a narrow financial task, but how increasingly agentic systems alter the structure of financial decision-making itself.图 1 展示了这一进程。该图是刻意风格化的：在实践中，这三个世代是重叠的，机构往往在同一工作流中结合使用它们。现代金融系统可能同时依赖算法执行、机器学习预测和代理推理。尽管如此，世代划分在分析上是有用的，因为它突显了自动化重心发生的真实转变。核心问题不再是人工智能能否辅助狭窄的金融任务，而是日益代理化的系统如何改变金融决策本身的结构。

This conceptual distinction motivates the next section. If agentic finance represents a shift toward workflow-centric automation, then the first task is to specify the internal structure of such systems. The following section therefore develops a modular architecture of AI financial agents and identifies the components through which they perceive information, reason under constraints, generate financial decision objects, and interact with execution and control infrastructures.这一概念性区分引出了下一节。如果代理金融代表了向以工作流为中心的自动化的转变，那么首要任务就是明确此类系统的内部结构。因此，下一节将开发金融人工智能代理的模块化架构，并识别它们感知信息、在约束下推理、生成金融决策对象以及与执行和控制基础设施交互的组件。

The first layer concerns how financial agents perceive and organise information. Financial environments are characterised by unusually heterogeneous data structures. Relevant signals may include high-frequency market prices, order-book dynamics, company disclosures, regulatory filings, earnings-call transcripts, macroeconomic releases, social-media signals, blockchain transaction flows, and internal portfolio data.第一层涉及金融代理如何感知和组织信息。金融环境的特征是异常异构的数据结构。相关信号可能包括高频市场价格、订单簿动态、公司披露、监管申报、财报电话会议记录、宏观经济发布、社交媒体信号、区块链交易流以及内部投资组合数据。

The primary challenge at this stage is not only data ingestion but also data governance. Financial information differs in latency, reliability, legal access rights, and economic interpretation. A market data feed may update thousands of times per second, whereas accounting information may change quarterly. Similarly, some information sources are public, while others are proprietary or subject to regulatory restrictions.这一阶段的主要挑战不仅在于数据摄入，还在于数据治理。金融信息在延迟、可靠性、合法访问权和经济解读方面存在差异。市场数据源每秒可能更新数千次，而会计信息可能每季度更新一次。同样，一些信息源是公开的，而另一些则是专有的或受监管限制的。

Consequently, the perception layer must perform several tasks simultaneously: data normalisation, timestamp alignment, provenance tracking, and access control. These functions determine which signals an agent can observe and how quickly it can respond to them. In practice, the design of this layer strongly shapes the informational advantage of financial agents, because differences in data pipelines often translate directly into differences in market responsiveness.因此，感知层必须同时执行多项任务：数据标准化、时间戳对齐、来源追踪和访问控制。这些功能决定了代理可以观察到哪些信号以及它能以多快的速度响应它们。在实践中，该层的设计极大地塑造了金融代理的信息优势，因为数据流水线的差异通常直接转化为市场响应能力的差异。

The second layer transforms perceived information into intermediate beliefs, forecasts, and explanations. This layer constitutes the cognitive core of the agent system.第二层将感知到的信息转化为中间信念、预测和解释。这一层构成了代理系统的认知核心。

In modern financial AI systems, the reasoning layer is rarely implemented as a single model. Instead, it typically combines multiple analytical components. Large language models provide capabilities for interpreting unstructured text such as filings, news, and regulatory documents. Retrieval systems enable agents to access historical context or domain knowledge. Statistical models and machine-learning algorithms generate forecasts of returns, volatility, credit risk, or liquidity conditions. Optimisation modules translate these forecasts into feasible portfolio or trading configurations.在现代金融人工智能系统中，推理层很少作为单一模型实现。相反，它通常结合了多种分析组件。大语言模型提供了解读非结构化文本（如申报文件、新闻和监管文档）的能力；检索系统使代理能够获取历史背景或领域知识；统计模型和机器学习算法生成关于收益、波动性、信用风险或流动性状况的预测；优化模块将这些预测转化为可行的投资组合或交易配置。

Recent developments in language-model agents further extend these capabilities by enabling sequential reasoning, planning, and tool usage (Yao et al., 2023; Park et al., 2023; Xi et al., 2023). In such systems, the reasoning engine can break complex tasks into intermediate steps, query external data sources, evaluate alternative hypotheses, and maintain internal memory across interactions. For financial decision-making, this hybrid design is particularly important because no single model type is sufficient to address the full range of analytical tasks involved.语言模型代理的最新发展通过实现序列推理、规划和工具使用，进一步扩展了这些能力（Yao et al., 2023; Park et al., 2023; Xi et al., 2023）。在此类系统中，推理引擎可以将复杂任务分解为中间步骤，查询外部数据源，评估替代假设，并在交互过程中保持内部记忆。对于金融决策而言，这种混合设计尤为重要，因为没有单一的模型类型足以处理所涉及的全部分析任务。

The third layer converts analytical outputs into decision objects. A decision object is a structured representation of a potential financial action together with the constraints and reasoning that justify it. Examples include trade proposals, portfolio reallocations, hedging strategies, anomaly alerts, compliance flags, or monitoring signals.第三层将分析输出转换为决策对象。决策对象是对潜在金融行为的结构化表示，包含证明其合理性的约束条件和推理逻辑。例如交易提案、投资组合再平衡、对冲策略、异常警报、合规标记或监控信号等。

This concept is important because it clarifies the role of AI agents in financial workflows. Rather than producing isolated predictions, agent systems generate decision objects that integrate information, reasoning, and institutional constraints. For example, a trading agent may produce a trade idea together with a confidence score, risk estimate, liquidity assessment, and explanation derived from recent news or macro events.这一概念之所以重要，是因为它阐明了人工智能代理在金融工作流中的角色。代理系统产生的不是孤立的预测，而是集成了信息、推理和机构约束的决策对象。例如，交易代理可能会生成一个交易想法，并附带置信度评分、风险评估、流动性评估以及基于近期新闻或宏观事件的解释。

Strategy generation is therefore the stage at which analytical insight becomes economically meaningful. At this stage, the agent must reconcile forecasts with practical constraints such as portfolio mandates, liquidity limits, regulatory requirements, and transaction costs. The design of this layer determines whether the agent functions primarily as a research assistant, a portfolio co-pilot, or a more autonomous decision system.因此，策略生成阶段是分析洞察转化为经济价值的阶段。在此阶段，代理必须协调预测与实际约束，如投资组合授权、流动性限制、监管要求和交易成本。该层的设计决定了代理主要充当研究助手、投资组合副驾驶，还是更具自主性的决策系统。

The final layer connects decision objects to financial infrastructure. This layer includes order-management systems, execution-management systems, exchange APIs, smart-contract interactions, and other operational interfaces through which financial actions are implemented.最后一层将决策对象连接到金融基础设施。该层包括订单管理系统、执行管理系统、交易所 API、智能合约交互以及实现金融行为的其他操作接口。

However, execution is not purely technical. It is also the point at which institutional control becomes critical. Financial institutions must ensure that AI-driven decisions remain consistent with risk limits, legal obligations, and governance frameworks. As a result, the execution layer typically incorporates approval workflows, position limits, audit trails, monitoring systems, and emergency stop mechanisms.然而，执行不仅仅是技术问题，也是机构控制至关重要的环节。金融机构必须确保人工智能驱动的决策与风险限额、法律义务和治理框架保持一致。因此，执行层通常包含审批工作流、头寸限额、审计追踪、监控系统和紧急停止机制。

The presence of these controls means that most real-world financial agents operate under bounded autonomy. Agents may generate recommendations or initiate actions within predefined limits, but human oversight remains responsible for escalation decisions and exceptional situations. From a systemic perspective, the design of the execution layer is particularly important because it determines how quickly agent decisions propagate into market activity.这些控制措施的存在意味着大多数现实世界的金融代理都在有限的自主权下运行。代理可以在预定义的限额内生成建议或发起行动，但对于升级决策和特殊情况，仍由人工监督负责。从系统角度来看，执行层的设计尤为重要，因为它决定了代理决策向市场活动传导的速度。

While the four-layer architecture describes the internal structure of financial agents, the broader theoretical implications of agentic finance depend on several design dimensions that cut across these layers. These dimensions serve as the key primitives in the Agentic Financial Market Model (AFMM) developed later in the paper.虽然四层架构描述了金融代理的内部结构，但代理金融更广泛的理论意义取决于贯穿这些层的若干设计维度。这些维度构成了本文后续开发的代理金融市场模型 (AFMM) 的关键原语。

• •

  Autonomy depth. The extent to which agents can act without human approval. Some systems only generate analytical summaries, while others can initiate trades or rebalance portfolios within predefined constraints.自主深度。 代理在无需人工批准的情况下采取行动的程度。有些系统仅生成分析摘要，而另一些系统可以在预定义的约束条件下发起交易或重新平衡投资组合。

• •

  Model heterogeneity. The diversity of models, prompts, and data sources used by different agents. Higher heterogeneity may increase informational diversity, while strong homogeneity may create correlated behaviour across institutions.模型异质性。 不同代理所使用的模型、提示词和数据源的多样性。较高的异质性可能会增加信息多样性，而强同质性可能会导致各机构间出现相关行为。

• •

  Execution coupling. The degree to which multiple agents respond to similar signals on similar timescales. Strong coupling may increase the speed of price discovery but can also amplify market volatility.执行耦合度。 多个代理在相似时间尺度上对相似信号做出反应的程度。强耦合可能会提高价格发现的速度，但也可能放大市场波动。

• •

  Infrastructure concentration. The extent to which agents rely on shared cloud platforms, model providers, APIs, or data vendors. High concentration introduces potential systemic dependencies.基础设施集中度。 代理对共享云平台、模型提供商、API 或数据供应商的依赖程度。高度集中会带来潜在的系统性依赖。

• •

  Supervisory observability. The ability of internal control teams or regulators to reconstruct the decision path of AI agents, including inputs, intermediate reasoning steps, and executed actions.监管可观测性。 内部控制团队或监管机构重建人工智能代理决策路径的能力，包括输入、中间推理步骤和已执行的操作。

Together, these dimensions connect micro-level system design to macro-level financial outcomes. They provide the analytical bridge between the architecture described in this section and the market-level framework introduced in the next section.总之，这些维度将微观层面的系统设计与宏观层面的金融结果联系起来。它们为本节描述的架构与下一节引入的市场层面框架之间提供了分析桥梁。

The central intuition of the AFMM is that the market impact of AI agents depends less on their individual intelligence and more on how they are distributed, coupled, and governed across the financial system.AFMM 的核心直觉在于，人工智能代理的市场影响与其个体智能关系不大，更多取决于它们在金融系统中的分布、耦合和治理方式。

A market populated by heterogeneous agents drawing on diverse models, data sources, and objectives may enhance price discovery by increasing informational diversity. In contrast, a market in which many institutions rely on highly similar agents, trained on overlapping datasets and connected to the same technological infrastructure, may produce correlated responses to common signals. Under such conditions, even locally optimal decisions can aggregate into market-wide instability.一个充斥着利用多样化模型、数据源和目标的异质性代理的市场，可能会通过增加信息多样性来增强价格发现。相反，一个许多机构依赖高度相似的代理（基于重叠数据集训练并连接到相同技术基础设施）的市场，可能会对共同信号产生相关反应。在这种条件下，即使是局部最优的决策也可能聚合成全市场的失稳。

The AFMM therefore shifts the focus from individual model performance to the population structure of agents. In this perspective, financial markets increasingly resemble ecosystems of interacting human and machine decision-makers embedded in shared infrastructures.因此，AFMM 将焦点从个体模型性能转移到了代理的种群结构上。从这个角度来看，金融市场越来越像是由嵌入在共享基础设施中的人类和机器决策者组成的交互生态系统。

Figure 3 summarises the conceptual structure of the AFMM. Agent design parameters determine how agents process information and interact in markets, which in turn shapes aggregate market outcomes.图 3 总结了 AFMM 的概念结构。代理设计参数决定了代理如何处理信息并在市场中进行交互，进而塑造了总体市场结果。

Consider a market with $N$ participants indexed by $i=1,\dots,N$. A subset of these participants employ AI agents as part of their financial decision workflow. For each AI-enabled participant, define the following design variables:考虑一个拥有 N 个参与者的市场，索引为 i=1,…,N。这些参与者中的一部分在其金融决策工作流中使用人工智能代理。对于每个启用人工智能的参与者，定义以下设计变量：

These variables correspond directly to the design dimensions introduced in the architecture framework of Section 4.这些变量直接对应于第 4 节架构框架中引入的设计维度。

• •

  Autonomy depth ($A_{i}$). The degree to which an agent can initiate or execute actions without human intervention.自主深度 (Ai)。 代理在无需人工干预的情况下启动或执行操作的程度。

• •

  Model heterogeneity ($H_{i}$). The diversity of models, data pipelines, prompts, and objectives across agents.模型异质性 (Hi)。 代理间在模型、数据管道、提示词和目标方面的多样性。

• •

  Execution coupling ($C_{i}$). The extent to which agents respond to similar signals on similar timescales.执行耦合度 (Ci)。 代理在相似时间尺度上对相似信号做出反应的程度。

• •

  Infrastructure concentration ($V_{i}$). The dependence of agents on common upstream providers such as cloud platforms, model vendors, data feeds, or middleware.基础设施集中度 (Vi)。 代理对云平台、模型供应商、数据源或中间件等共同上游提供商的依赖。

• •

  Supervisory observability ($S_{i}$). The ability of institutions or regulators to reconstruct and intervene in the decision paths of agents.监管可观测性 (Si)。 机构或监管机构重建和干预代理决策路径的能力。

Let market-level outcomes be represented by the following variables:令市场层面的结果由以下变量表示：

These outcomes capture the primary dimensions along which agentic finance may influence market structure.这些结果捕捉了代理金融可能影响市场结构的主要维度。

To connect agent design parameters to market behaviour, we introduce a stylised representation of the decision process of AI-enabled financial agents.为了将代理设计参数与市场行为联系起来，我们引入了人工智能金融代理决策过程的风格化表示。

At each time $t$, agent $i$ observes a vector of market signals在每个时间 t，代理 i 观察到一个市场信号向量

where $P_{t}$ denotes asset prices, $Z_{t}$ represents macro or market state variables, and $N_{t}$ captures textual or informational signals such as news or disclosures.其中 Pt 表示资产价格，Zt 表示宏观或市场状态变量，Nt 捕捉新闻或披露等文本或信息信号。

Given these inputs, the agent produces a decision object $D_{i,t}$ through a decision function给定这些输入，代理通过决策函数产生一个决策对象 Di,t

where $g_{i}(\cdot)$ represents the internal reasoning pipeline of the agent. This function abstracts the layered architecture described earlier in the paper, combining data perception, reasoning, strategy generation, and execution constraints.其中 gi(·) 表示代理的内部推理管道。该函数抽象了本文前面描述的分层架构，结合了数据感知、推理、策略生成和执行约束。

The parameters $(A_{i},H_{i},C_{i},V_{i},S_{i})$ influence the properties of this decision rule. Higher autonomy depth ($A_{i}$) allows the agent to initiate actions with fewer human approval constraints. Greater heterogeneity ($H_{i}$) reflects diversity in models, training data, and objectives across agents. Execution coupling ($C_{i}$) captures the extent to which agents respond to common signals or triggers. Infrastructure concentration ($V_{i}$) represents reliance on shared technological providers, while supervisory observability ($S_{i}$) captures institutional control capacity and monitoring visibility.参数 (Ai, Hi, Ci, Vi, Si) 影响此决策规则的属性。较高的自主深度 (Ai) 允许代理在较少的人工审批约束下启动操作。更大的异质性 (Hi) 反映了代理间在模型、训练数据和目标方面的多样性。执行耦合度 (Ci) 捕捉了代理对共同信号或触发因素的反应程度。基础设施集中度 (Vi) 代表对共享技术提供商的依赖，而监管可观测性 (Si) 则捕捉机构的控制能力和监控可见性。

The output $D_{i,t}$ may correspond to various financial actions, including trade instructions, portfolio rebalancing proposals, risk alerts, or monitoring signals.输出 Di,t 可能对应各种金融行为，包括交易指令、投资组合再平衡建议、风险警报或监控信号。

Because financial agents do not always execute their recommendations directly, it is useful to distinguish between a decision object and its realised market action. Let the realised action be由于金融代理并不总是直接执行其建议，区分决策对象及其实现的市场行为非常有用。令已实现的行为为

where $\Lambda_{i}(\cdot)$ maps decision objects into executed actions through approval rules, limits, supervisory checks, and other institutional controls. This distinction is important because two agents may generate similar internal recommendations while differing sharply in the speed or scale with which those recommendations reach the market.其中 Λi(·) 通过审批规则、限额、监管检查和其他机构控制将决策对象映射为已执行的行为。这种区分很重要，因为两个代理可能生成相似的内部建议，但在这些建议到达市场的速度或规模上却大相径庭。

To capture the possibility of correlated machine-mediated behaviour, define the average similarity of realised actions as为了捕捉机器中介行为的相关性可能性，定义已实现行为的平均相似度为

where higher $\rho_{t}$ corresponds to greater synchronisation in realised actions. In the spirit of the AFMM, $\rho_{t}$ should be increasing in execution coupling and common infrastructure dependence, and decreasing in model heterogeneity. In reduced form, this implies其中较高的 ρt 对应于已实现行为的更高同步性。本着 AFMM 的精神，ρt 应随执行耦合度和共同基础设施依赖性的增加而增加，随模型异质性的增加而减少。简化形式下，这意味着

Market outcomes emerge from the aggregation of realised rather than merely recommended actions. The aggregate market action can therefore be written as市场结果源于已实现行为而非仅仅是建议行为的聚合。因此，总的市场行为可以写为

where $w_{i}$ represents the market weight or trading capacity of participant $i$.其中 wi 表示参与者 i 的市场权重或交易能力。

Asset prices and liquidity conditions evolve as a function of the aggregate order flow:资产价格和流动性状况随总订单流的函数而演变：

where $F(\cdot)$ captures the market impact of aggregated decisions under prevailing market conditions.其中 F(·) 捕捉了在当前市场条件下聚合决策的市场影响。

More generally, the market state evolves according to更一般地，市场状态根据以下方式演变

where $\varepsilon_{t+1}$ denotes exogenous news and liquidity shocks. The AFMM is therefore explicitly dynamic: agent design affects realised actions, realised actions affect market states, and updated market states feed back into future agent decisions.其中 εt+1 表示外生新闻和流动性冲击。因此，AFMM 是显式动态的：代理设计影响已实现的行为，已实现的行为影响市场状态，而更新后的市场状态又反馈到未来的代理决策中。

This representation highlights a key feature of agentic finance: market dynamics depend not only on individual decision quality but also on the correlation structure of agent decisions. When many agents rely on similar models, signals, or infrastructure, the resulting decisions may become highly correlated, amplifying market reactions.这种表示强调了代理金融的一个关键特征：市场动态不仅取决于个体决策质量，还取决于代理决策的相关性结构。当许多代理依赖相似的模型、信号或基础设施时，产生的决策可能会变得高度相关，从而放大市场反应。

In this sense, the AFMM focuses on how the joint distribution of agent design parameters $(A_{i},H_{i},C_{i},V_{i},S_{i})$ shapes the statistical properties of the aggregate decision process $Q_{t}$ through both the scale of realised actions and the correlation structure summarised by $\rho_{t}$.从这个意义上说，AFMM 侧重于代理设计参数 (Ai, Hi, Ci, Vi, Si) 的联合分布如何通过已实现行为的规模和由 ρt 总结的相关性结构来塑造聚合决策过程 Qt 的统计特性。

The AFMM assumes that market outcomes emerge from the interaction between the cross-sectional distribution of agent design parameters and the prevailing market environment.AFMM 假设市场结果源于代理设计参数的横截面分布与当前市场环境之间的相互作用。

In reduced form, market-level outcomes can be written as:简化形式下，市场层面的结果可以写为：

where bars denote market-level averages or concentration-weighted aggregates and $X$ represents exogenous market conditions such as macroeconomic shocks or liquidity stress.其中横线表示市场层面的平均值或集中度加权聚合，X 表示宏观经济冲击或流动性压力等外生市场条件。

The qualitative logic underlying these relationships can be summarised as follows.这些关系背后的定性逻辑可总结如下。

First, greater autonomy increases the speed and scale of market reactions. While this may enhance efficiency under normal conditions, it may also amplify rapid feedback loops during periods of stress.首先，更大的自主性增加了市场反应的速度和规模。虽然这在正常条件下可能会提高效率，但也可能在压力时期放大快速反馈循环。

Second, greater heterogeneity across models and data pipelines increases informational diversity. This tends to improve price discovery by ensuring that agents interpret signals differently rather than converging on a single dominant narrative.其次，模型和数据管道的更大异质性增加了信息多样性。这往往通过确保代理以不同的方式解读信号而非收敛于单一的主导叙事，从而改善价格发现。

Third, strong execution coupling increases the probability of synchronous reactions. When many agents respond to similar signals at similar times, market reactions may become highly correlated, pushing $\rho_{t}$ upward.第三，强执行耦合增加了同步反应的可能性。当许多代理在相似的时间对相似的信号做出反应时，市场反应可能会变得高度相关，从而推高 ρt。

Fourth, dependence on common technological infrastructure introduces common-mode failure risk. If many institutions rely on the same upstream providers, outages or model failures may propagate simultaneously across the system and increase the effective similarity of realised actions even when institutions appear organizationally distinct.第四，对共同技术基础设施的依赖引入了共模故障风险。如果许多机构依赖相同的上游提供商，中断或模型故障可能会在整个系统中同时传播，并增加已实现行为的有效相似性，即使这些机构在组织上看起来是截然不同的。

Finally, strong supervisory observability acts as a stabilising force by enabling effective monitoring, intervention, and governance.最后，强大的监管可观测性通过实现有效的监控、干预和治理，发挥了稳定力量的作用。

The AFMM generates several qualitative propositions about the behaviour of agentic financial markets.AFMM 产生了关于代理金融市场行为的几个定性命题。

Although the AFMM is primarily conceptual, its structure naturally suggests empirical hypotheses that can be tested using market data or simulation environments.尽管 AFMM 主要处于概念层面，但其结构自然地提出了可以使用市场数据或模拟环境进行测试的实证假设。

The AFMM strengthens the theoretical contribution of this paper in two ways. First, it transforms the architectural framework of Section 4 into a mechanism-based explanation of market dynamics. Rather than treating AI adoption as a purely technological phenomenon, the model highlights the systemic consequences of agent design choices.AFMM 通过两种方式加强了本文的理论贡献。首先，它将第 4 节的架构框架转化为基于机制的市场动态解释。该模型并未将人工智能采用视为纯粹的技术现象，而是强调了代理设计选择的系统性后果。

Second, the AFMM creates a bridge between conceptual analysis and empirical research. By expressing market outcomes as functions of agent design parameters, the framework provides a structure within which future empirical studies can test how different configurations of agentic finance influence market efficiency, volatility, and systemic stability.其次，AFMM 在概念分析与实证研究之间架起了一座桥梁。通过将市场结果表示为代理设计参数的函数，该框架提供了一个结构，未来的实证研究可以在此结构内测试代理金融的不同配置如何影响市场效率、波动性和系统稳定性。

Trading is the most natural domain for the deployment of AI agents because it already relies heavily on automated infrastructure such as algorithmic execution systems, market data pipelines, and order management platforms. However, the transition from algorithmic trading to agentic trading primarily concerns workflow integration rather than complete autonomy.交易是部署人工智能代理最自然的领域，因为它已经严重依赖算法执行系统、市场数据管道和订单管理平台等自动化基础设施。然而，从算法交易向代理交易的转型主要涉及工作流集成，而非完全自主。

In traditional algorithmic trading systems, strategies are typically predefined and executed through rule-based engines. AI agents extend this architecture by integrating multiple tasks within a single decision pipeline. For example, an agent may simultaneously monitor news flows, interpret macroeconomic signals, update forecasts, and generate candidate trade instructions. The agent can then pass proposed orders through execution algorithms subject to risk limits and compliance filters.在传统的算法交易系统中，策略通常是预定义的，并通过基于规则的引擎执行。人工智能代理通过在单一决策管道内集成多项任务来扩展此架构。例如，代理可以同时监控新闻流、解读宏观经济信号、更新预测并生成候选交易指令。然后，代理可以将提议的订单通过受风险限额和合规过滤器约束的执行算法。

In practice, most institutions are likely to adopt bounded autonomy models in which AI agents assist with signal synthesis, order preparation, and post-trade analysis rather than initiating large-scale autonomous trading. Such designs reduce operational risk while still allowing agents to improve reaction speed and information processing capacity.在实践中，大多数机构可能会采用有限自主模型，即人工智能代理协助信号合成、订单准备和交易后分析，而不是发起大规模的自主交易。这种设计降低了操作风险，同时仍允许代理提高反应速度和信息处理能力。

Recent developments in LLM-based trading frameworks illustrate this shift toward agentic workflows. Multi-agent trading environments allow different agents to specialise in forecasting, portfolio construction, and risk monitoring, while interacting through structured communication protocols (Yang et al., 2023; Xiao et al., 2025). These architectures resemble distributed research teams rather than single autonomous traders.基于大语言模型的交易框架的最新发展说明了这种向代理工作流的转变。多代理交易环境允许不同的代理专注于预测、投资组合构建和风险监控，同时通过结构化的通信协议进行交互（Yang et al., 2023; Xiao et al., 2025）。这些架构更像是分布式的研究团队，而非单一的自主交易者。

Within the AFMM framework, trading applications are particularly sensitive to the parameters of autonomy depth ($A$) and execution coupling ($C$). High autonomy combined with strong signal coupling could amplify synchronous reactions across institutions, potentially contributing to episodic volatility events similar to those observed in earlier algorithmic trading episodes (Kirilenko et al., 2017).在 AFMM 框架内，交易应用对自主深度 (AA) 和执行耦合度 (CC) 的参数特别敏感。高自主性与强信号耦合相结合可能会放大机构间的同步反应，可能导致类似于早期算法交易事件中观察到的偶发性波动事件（Kirilenko et al., 2017）。

In portfolio management, the most plausible role for AI agents is not full automation but research augmentation. Asset allocation decisions involve complex trade-offs between macroeconomic expectations, valuation models, institutional mandates, and client preferences. These tasks require structured reasoning, scenario construction, and narrative explanation—areas where large language models and agentic workflows can provide meaningful assistance.在投资组合管理中，人工智能代理最合理的角色不是完全自动化，而是研究增强。资产配置决策涉及宏观经济预期、估值模型、机构授权和客户偏好之间的复杂权衡。这些任务需要结构化推理、情景构建和叙事解释——而这正是大语言模型和代理工作流可以提供有意义帮助的领域。

AI agents can support portfolio managers by synthesising large volumes of heterogeneous information, including macroeconomic releases, corporate disclosures, earnings calls, and geopolitical developments. Instead of producing simple predictions, agents can generate structured research outputs such as investment memos, scenario analyses, and portfolio stress tests.人工智能代理可以通过合成海量异质性信息（包括宏观经济发布、公司披露、财报电话会议和地缘政治发展）来支持投资组合经理。代理可以生成结构化的研究输出（如投资备忘录、情景分析和投资组合压力测试），而不是仅仅产生简单的预测。

Another promising application lies in scenario generation and probabilistic reasoning. Agents can construct multiple hypothetical macroeconomic trajectories and evaluate how different asset classes may perform under alternative policy or market conditions. Such capabilities align closely with the reasoning and planning modules typical of agentic architectures.另一个有前景的应用在于情景生成和概率推理。代理可以构建多种假设的宏观经济轨迹，并评估不同资产类别在替代政策或市场条件下的表现。这些能力与代理架构中典型的推理和规划模块紧密结合。

Importantly, the fiduciary nature of asset management imposes strong constraints on autonomy depth. Most portfolio decisions require explicit human approval, meaning that agents primarily function as co-pilot systems rather than independent allocators. In the AFMM framework, this corresponds to relatively low values of $A$ but potentially high values of informational heterogeneity ($H$), which may enhance the diversity of analytical perspectives available to portfolio managers.重要的是，资产管理的受托性质对自主深度施加了强约束。大多数投资组合决策需要明确的人工批准，这意味着代理主要充当副驾驶系统，而非独立的分配者。在 AFMM 框架中，这对应于相对较低的 AA 值，但可能具有较高的信息异质性 (HH) 值，这可能会增加投资组合经理可获得的分析视角的广度。

Risk management and compliance monitoring represent one of the most immediately impactful domains for AI agents in finance. Financial institutions already operate extensive monitoring systems for market risk, credit exposure, operational risk, and regulatory compliance. However, these systems often rely on static rule sets or manually maintained indicators.风险管理和合规监控代表了人工智能代理在金融领域最直接产生影响的领域之一。金融机构已经为市场风险、信用敞口、操作风险和监管合规运行了广泛的监控系统。然而，这些系统通常依赖静态规则集或手动维护的指标。

Agentic systems can enhance these monitoring frameworks by continuously analysing transaction flows, market data, and internal control signals in real time. AI agents can detect anomalous patterns, flag suspicious transactions, or identify emerging liquidity stress before traditional indicators trigger alerts.代理系统可以通过实时持续分析交易流、市场数据和内部控制信号来增强这些监控框架。人工智能代理可以检测异常模式、标记可疑交易或在传统指标触发警报之前识别新兴的流动性压力。

From a regulatory perspective, the ability of AI agents to process large volumes of heterogeneous information may improve the timeliness and granularity of risk detection (Aldasoro et al., 2024). For example, agents could monitor market microstructure indicators, news flows, and funding conditions simultaneously to identify early signs of systemic stress.从监管角度来看，人工智能代理处理海量异质性信息的能力可能会提高风险检测的及时性和粒度（Aldasoro et al., 2024）。例如，代理可以同时监控市场微观结构指标、新闻流和融资条件，以识别系统性压力的早期迹象。

However, the deployment of AI agents in compliance settings raises important governance challenges. False positives may generate excessive alerts, while opaque decision processes may complicate accountability. These concerns highlight the importance of supervisory observability ($S$) within the AFMM. Institutions must ensure that AI-generated alerts can be audited, interpreted, and challenged by human supervisors.然而，在合规环境中部署人工智能代理引发了重要的治理挑战。误报可能会产生过多的警报，而不透明的决策过程可能会使问责变得复杂。这些担忧凸显了 AFMM 中监管可观测性 (SS) 的重要性。机构必须确保人工智能生成的警报可以由人工主管进行审计、解释和质疑。

More broadly, AI agents may play an important role in the supervision of financial markets themselves. Regulators are increasingly exploring AI tools to monitor trading behaviour, detect market manipulation, and analyse systemic risk indicators across multiple institutions (Crisanto et al., 2024; Aquilina et al., 2025).更广泛地说，人工智能代理可能在金融市场本身的监管中发挥重要作用。监管机构正越来越多地探索人工智能工具，以监控交易行为、检测市场操纵并分析跨机构的系统性风险指标（Crisanto et al., 2024; Aquilina et al., 2025）。

Decentralised finance (DeFi) ecosystems provide a particularly fertile environment for AI agents because financial activity on blockchains generates large volumes of publicly observable data. Transactions, liquidity flows, governance proposals, and smart contract interactions are recorded on-chain, creating rich datasets for automated analysis.去中心化金融 (DeFi) 生态系统为人工智能代理提供了一个特别肥沃的环境，因为区块链上的金融活动会产生大量可公开观察的数据。交易、流动性流、治理提案和智能合约交互都被记录在链上，为自动化分析创造了丰富的数据集。

AI agents can analyse blockchain data to trace wallet activity, detect unusual trading behaviour, or monitor governance dynamics in decentralised protocols. For example, agents may track large liquidity movements across decentralised exchanges, identify emerging arbitrage opportunities, or evaluate the security risks of newly deployed smart contracts.人工智能代理可以分析区块链数据以追踪钱包活动、检测异常交易行为或监控去中心化协议中的治理动态。例如，代理可以追踪去中心化交易所之间的大额流动性变动、识别新兴套利机会或评估新部署智能合约的安全风险。

Another important application lies in the monitoring of decentralised governance processes. AI agents can analyse proposal discussions, voting behaviour, and token distributions to assess the likelihood of governance capture or coordinated manipulation.另一个重要的应用在于去中心化治理过程的监控。人工智能代理可以分析提案讨论、投票行为和代币分布，以评估治理捕获或协调操纵的可能性。

The integration of AI agents with DeFi protocols also enables new forms of automated financial services. Agents may dynamically allocate liquidity across decentralised lending platforms, manage collateral positions, or monitor liquidation risks in real time. Because DeFi protocols operate continuously and transparently, such agents can interact directly with smart contracts without the operational frictions typical of traditional financial infrastructure.人工智能代理与 DeFi 协议的集成也实现了新型的自动化金融服务。代理可以动态地在去中心化借贷平台间分配流动性、管理抵押头寸或实时监控清算风险。由于 DeFi 协议持续且透明地运行，此类代理可以直接与智能合约交互，而无需传统金融基础设施中典型的操作摩擦。

Within the AFMM framework, DeFi applications illustrate the interaction between high autonomy depth ($A$) and fragmented infrastructure. While blockchain transparency increases observability, fragmented liquidity across multiple protocols may create new forms of execution coupling and systemic fragility.在 AFMM 框架内，DeFi 应用说明了高自主深度 (AA) 与碎片化基础设施之间的相互作用。虽然区块链的透明度提高了可观测性，但跨多个协议的碎片化流动性可能会产生新型的执行耦合和系统性脆弱性。

Table 1 summarises the main application domains of AI agents in finance, highlighting typical tasks, autonomy levels, and institutional constraints. Together, these examples illustrate that the most significant impact of agentic finance may arise not from fully autonomous decision systems but from the gradual integration of AI agents into existing financial workflows.表 1 总结了人工智能代理在金融中的主要应用领域，强调了典型任务、自主水平和机构约束。总之，这些例子说明，代理金融最重大的影响可能并非来自完全自主的决策系统，而是来自人工智能代理向现有金融工作流的逐步集成。

Across these domains, a common pattern emerges. The immediate economic impact of agentic finance is likely to come less from fully autonomous market actors than from the gradual reconfiguration of research, monitoring, execution, and control workflows inside existing institutions. That pattern is precisely why governance becomes central. Once agentic systems are embedded in material financial processes, the key question is no longer whether they improve local tasks, but whether their delegation boundaries, infrastructure dependencies, and oversight mechanisms are robust enough for system-wide adoption.在这些领域中，出现了一种共同模式。代理金融的直接经济影响可能更多地来自现有机构内部研究、监控、执行和控制工作流的逐步重构，而非完全自主的市场参与者。这种模式正是治理变得至关重要的原因。一旦代理系统嵌入到实质性的金融过程中，关键问题就不再是它们是否改进了局部任务，而是它们的委托边界、基础设施依赖性和监督机制是否足以应对全系统的采用。

The increasing deployment of AI agents introduces several governance challenges for financial institutions and regulators.人工智能代理的日益普及为金融机构和监管机构带来了多项治理挑战。

First, the use of autonomous or semi-autonomous agents complicates the attribution of responsibility. When decisions emerge from interactions among multiple models, data sources, and optimisation modules, identifying the origin of an error or unexpected market outcome becomes more difficult. This raises important questions regarding accountability, auditability, and model governance.首先，自主或半自主代理的使用使责任归属变得复杂。当决策源于多个模型、数据源和优化模块之间的交互时，确定错误或意外市场结果的来源变得更加困难。这引发了关于问责制、可审计性和模型治理的重要问题。

Second, AI agents may increase the opacity of financial decision processes. Complex model pipelines and adaptive learning systems can make it difficult for supervisors to reconstruct the reasoning behind specific decisions. Regulatory frameworks therefore increasingly emphasise explainability and traceability in AI-enabled financial systems (Financial Stability Board, 2024).其次，人工智能代理可能会增加金融决策过程的不透明度。复杂的模型管道和自适应学习系统可能会使监管机构难以重建特定决策背后的推理。因此，监管框架正日益强调人工智能驱动的金融系统中的可解释性和可追溯性（金融稳定委员会，2024）。

Third, the concentration of technological infrastructure introduces new forms of systemic risk. Many financial institutions rely on common upstream providers for cloud computing, foundation models, data services, or trading infrastructure. If a large number of AI agents depend on similar models or APIs, failures in these shared components may propagate across the financial system.第三，技术基础设施的集中化引入了新型的系统性风险。许多金融机构在云计算、基础模型、数据服务或交易基础设施方面依赖共同的上游提供商。如果大量人工智能代理依赖相似的模型或 API，这些共享组件中的故障可能会在整个金融系统中传播。

Fourth, AI agents may accelerate the speed of market reactions. While faster information processing can improve efficiency, it may also amplify feedback loops during periods of stress. These dynamics resemble earlier concerns about high-frequency trading, but may operate at a broader scale because agentic systems integrate information processing, reasoning, and decision-making in a single pipeline.第四，人工智能代理可能会加快市场反应的速度。虽然更快的信息处理可以提高效率，但也可能在压力时期放大反馈循环。这些动态类似于早期关于高频交易的担忧，但由于代理系统将信息处理、推理和决策集成在单一管道中，其影响范围可能更广。

Within the AFMM framework, these governance issues correspond directly to the model parameters of infrastructure concentration ($V$), supervisory observability ($S$), and execution coupling ($C$). Effective governance therefore requires institutional mechanisms that maintain sufficient observability and control as agent autonomy increases.在 AFMM 框架内，这些治理问题直接对应于基础设施集中度 (VV)、监管可观测性 (SS) 和执行耦合度 (CC) 的模型参数。因此，有效的治理需要机构机制，在代理自主性增加的同时保持足够的可观测性和控制力。

Regulators and international organisations have begun to address the implications of AI adoption in financial markets. Recent policy discussions emphasise several supervisory strategies.监管机构和国际组织已开始着手解决人工智能在金融市场中应用的影响。近期的政策讨论强调了若干监管策略。

First, institutions are expected to implement robust model governance frameworks that document training data, model architecture, and decision logic. Such documentation helps ensure that AI-driven decisions remain auditable and accountable.首先，机构被期望实施稳健的模型治理框架，记录训练数据、模型架构和决策逻辑。此类文档有助于确保人工智能驱动的决策保持可审计和可问责。

Second, supervisory authorities may require financial firms to maintain human-in-the-loop controls for high-impact financial decisions. In most jurisdictions, full delegation of fiduciary decisions to autonomous AI systems remains unlikely in the near term.其次，监管机构可能要求金融公司对高影响力的金融决策维持“人在回路”的控制。在大多数司法管辖区，在短期内将受托决策完全委托给自主人工智能系统仍然不太可能。

Third, regulators increasingly monitor technological concentration risks arising from shared infrastructure. If many institutions rely on the same cloud providers, model vendors, or data pipelines, operational failures could generate correlated disruptions across the financial system.第三，监管机构正日益监控由共享基础设施引起的技术集中风险。如果许多机构依赖相同的云提供商、模型供应商或数据管道，操作故障可能会在整个金融系统中产生相关的干扰。

Fourth, AI systems themselves may become tools for financial supervision. Regulatory authorities are exploring AI-driven monitoring systems capable of analysing market data, transaction flows, and textual information to detect emerging risks in real time (Aldasoro et al., 2024; Aquilina et al., 2025).第四，人工智能系统本身可能成为金融监管的工具。监管机构正在探索能够分析市场数据、交易流和文本信息以实时检测新兴风险的人工智能驱动监控系统（Aldasoro et al., 2024; Aquilina et al., 2025）。

These developments suggest that financial stability outcomes will depend not only on technological adoption but also on the co-evolution of institutional governance and supervisory oversight.这些发展表明，金融稳定结果不仅取决于技术采用，还取决于机构治理与监管监督的共同演进。

The AFMM framework implies a set of governance principles that can guide the safe deployment of AI agents in financial markets.AFMM 框架提出了一套治理原则，可为 AI 智能体在金融市场中的安全部署提供指导。

1. P1:

   Bounded autonomy. Delegation to AI agents should depend on the materiality, reversibility, and legal consequences of decisions. High-impact financial actions should remain subject to human approval.有限自主权。 对 AI 智能体的授权应取决于决策的重要性、可逆性以及法律后果。高影响力的金融行动应始终需要人工审批。

2. P2:

   Traceability and auditability. AI-generated outputs should be linked to verifiable records of data sources, model versions, prompts, and execution instructions. Such traceability is essential for accountability and regulatory review.可追溯性与可审计性。 AI 生成的输出应与可验证的数据源、模型版本、提示词和执行指令记录相关联。这种可追溯性对于问责制和监管审查至关重要。

3. P3:

   Diversity by design. Institutions should actively monitor concentration in models, datasets, and technological infrastructure. Maintaining diversity in analytical approaches reduces the risk of correlated errors.设计层面的多样性。 机构应积极监测模型、数据集和技术基础设施的集中度。保持分析方法的多样性可降低相关性错误带来的风险。

4. P4:

   Embedded intervention capacity. Agentic systems should incorporate built-in control mechanisms such as approval gates, exposure limits, and emergency shutdown procedures.内置干预能力。 智能体系统应纳入内置的控制机制，例如审批门控、风险敞口限制和紧急关停程序。

5. P5:

   Supervisory co-evolution. Regulators may need to deploy AI-enabled monitoring systems of their own. As financial institutions adopt agentic systems, supervisory technologies must evolve to maintain effective oversight.监管协同演进。 监管机构可能需要部署其自身的 AI 赋能监测系统。随着金融机构采用智能体系统，监管技术必须不断演进，以维持有效的监督。

These principles follow directly from the AFMM. When autonomy depth ($A$), execution coupling ($C$), and infrastructure concentration ($V$) increase, supervisory observability ($S$) must increase correspondingly to preserve systemic stability.这些原则直接源自 AFMM。当自主深度 (AA)、执行耦合度 (CC) 和基础设施集中度 (VV) 增加时，监管可观测性 (SS) 也必须相应增加，以维护系统稳定性。

The next section takes this argument one step further by asking how the agentic-finance framework can be examined empirically. Rather than attempting a full structural test, it develops a compact event-study design that links public capability disclosures to market repricing.下一节将探讨如何对智能体金融框架进行实证检验，从而将这一论点推进一步。本研究并未尝试进行全面的结构性测试，而是开发了一种紧凑的事件研究设计，将公开的能力披露与市场定价调整联系起来。

This section develops an illustrative empirical application focused on AI-agent capability shocks: public disclosures that shift market beliefs about the scope, speed, and workflow relevance of agentic AI. The mechanism of interest is an expectations channel. When a credible capability disclosure suggests that agentic systems can accelerate legacy-code renovation, automate parts of software modernization, or reduce the cost of technical migration, investors should update expected future cash flows differently across firms.本节开发了一个说明性的实证应用，重点关注 AI 智能体能力冲击：即那些改变市场对智能体 AI 的范围、速度和工作流相关性预期的公开披露。我们感兴趣的机制是预期渠道。当可信的能力披露表明智能体系统可以加速遗留代码更新、自动化部分软件现代化或降低技术迁移成本时，投资者应相应地更新对不同公司未来现金流的预期。

This mechanism is tightly connected to the central theme of the paper. The architecture discussion in Section 4 argues that agentic systems matter because they combine perception, reasoning, planning, and tool use within end-to-end workflows. The applications discussion in Section 6 shows that such systems can augment or substitute for specific financial and operational tasks. The systemic question is therefore not only whether model quality improves, but which business models become more valuable or more vulnerable when agentic capability becomes more credible. The application examines that question through secondary-market prices.这一机制与本文的核心主题紧密相连。第 4 节关于架构的讨论指出，智能体系统之所以重要，是因为它们在端到端工作流中结合了感知、推理、规划和工具使用。第 6 节关于应用的讨论表明，此类系统可以增强或替代特定的金融和运营任务。因此，系统性问题不仅在于模型质量是否提高，还在于当智能体能力变得更加可信时，哪些商业模式会变得更有价值或更脆弱。本应用通过二级市场价格来审视这一问题。

Early 2026 provides a tractable setting. Public materials from Anthropic highlighted agentic code-modernisation workflows, especially COBOL migration and legacy-system renovation use cases (Anthropic, 2025; Anthropic, 2026b; Anthropic, 2026a). Anthropic’s March 5, 2026 research note then introduced the idea of observed exposure, explicitly distinguishing realised usage from theoretical capability and showing that market expectations may move faster than observable labour-market outcomes (Massenkoff and McCrory, 2026). Contemporaneous reporting further suggests that these disclosures were associated with sharp repricing in firms connected to legacy-code modernisation (Reuters, 2026). For the purposes of this paper, the importance of these events is not that they prove widespread deployment. Rather, they provide a clean setting in which to study how capability news is transmitted into market valuations. In the event calendar used for the empirical application, the February 23, 2026 CNBC coverage is retained as a media-amplification date, but it is not tabulated separately because it overlaps tightly with the February 24 market-validation window and would otherwise double-count the same information shock. A fuller implementation could nevertheless use February 23 as a robustness event or cluster it jointly with February 24.2026 年初提供了一个易于分析的环境。Anthropic 的公开材料强调了智能体代码现代化工作流，特别是 COBOL 迁移和遗留系统更新用例（Anthropic, 2025; Anthropic, 2026b; Anthropic, 2026a）。随后，Anthropic 2026 年 3 月 5 日的研究报告引入了“观测到的风险敞口”概念，明确区分了已实现的使用量与理论能力，并表明市场预期可能比可观察的劳动力市场结果移动得更快（Massenkoff and McCrory, 2026）。同期报道进一步表明，这些披露与涉及遗留代码现代化的公司股价剧烈变动有关（Reuters, 2026）。就本文而言，这些事件的重要性不在于证明了广泛的部署，而在于它们提供了一个清晰的环境，用于研究能力新闻如何转化为市场估值。在实证应用所使用的事件日历中，2026 年 2 月 23 日的 CNBC 报道被保留作为媒体放大日期，但未单独列出，因为它与 2 月 24 日的市场验证窗口紧密重叠，否则会导致对同一信息冲击的重复计算。尽管如此，更全面的实施可以将 2 月 23 日作为稳健性事件，或将其与 2 月 24 日合并在一起进行聚类分析。

The empirical question can be stated as follows:实证问题可表述如下：

How do public disclosures about agentic AI capabilities affect the valuation of firms with different exposures to legacy modernisation, workflow substitution, and financial-sector technology debt?关于智能体 AI 能力的公开披露，如何影响那些在遗留系统现代化、工作流替代和金融部门技术债务方面具有不同风险敞口的公司估值？

To answer this question, the sample is organised into three economically meaningful groups. The first group contains legacy-service vendors: firms whose revenues are plausibly linked to legacy maintenance, mainframe support, or large-scale modernization consulting. The second group contains legacy-heavy financial users: listed firms in banking, insurance, and payments whose operating environments may depend on older core systems. The third group is a control portfolio of large software and infrastructure firms that are adjacent to the same technology ecosystem but are not obviously centred on legacy-code dependence.为了回答这个问题，样本被组织为三个具有经济意义的组别。第一组包含遗留服务供应商：其收入与遗留维护、大型机支持或大规模现代化咨询有明确联系的公司。第二组包含重度依赖遗留系统的金融用户：银行、保险和支付领域的上市公司，其运营环境可能依赖于较旧的核心系统。第三组是大型软件和基础设施公司的对照组合，它们处于相同的技术生态系统中，但并非明显以遗留代码依赖为中心。

This taxonomy matters for the finance theme of the paper. In banking, insurance, and payments, COBOL and mainframe dependence remain economically relevant because core transaction systems, policy administration, payments processing, and settlement infrastructure are difficult to replace quickly. A capability shock that makes modernization more credible may therefore have opposite implications across firms. For incumbent service vendors, it may imply revenue pressure or margin compression. For legacy-heavy users, it may imply lower future migration costs, lower technology debt, or improved operating flexibility.这种分类对于本文的金融主题至关重要。在银行、保险和支付领域，对 COBOL 和大型机的依赖在经济上仍然具有相关性，因为核心交易系统、保单管理、支付处理和结算基础设施难以快速替换。因此，一个使现代化变得更加可信的能力冲击，可能会在不同公司之间产生相反的影响。对于在位服务供应商而言，这可能意味着收入压力或利润率压缩。对于重度依赖遗留系统的用户而言，这可能意味着更低的未来迁移成本、更少的技术债务或更高的运营灵活性。

The key firm-level explanatory variable is a filing-based legacy-modernization score constructed from SEC disclosures:关键的公司层面解释变量是根据 SEC 披露信息构建的基于申报文件的遗留现代化评分：

where $f_{ik}$ captures the frequency or salience of keyword $k$ in firm $i$’s 10-K or 10-Q filings. In the first-pass implementation, the keyword set includes terms related to the legacy core stack (COBOL, mainframe, z/OS, DB2, CICS), modernization processes (application modernization, cloud migration, replatforming, technical debt), and financial-system context (core banking, payment processing, policy administration). The baseline score uses simple term weights; a richer version could replace this with inverse-document-frequency or contextual NLP weights.其中 fi​kf_{ik} 捕获了关键字 kk 在公司 ii 的 10-K 或 10-Q 申报文件中的频率或显著性。在初步实施中，关键字集包括与遗留核心栈（COBOL、大型机、z/OS、DB2、CICS）、现代化流程（应用现代化、云迁移、平台重构、技术债务）和金融系统背景（核心银行、支付处理、保单管理）相关的术语。基准评分使用简单的词权重；更丰富的版本可以用逆文档频率或上下文 NLP 权重来替代。

The main outcome variables are cumulative abnormal returns and abnormal trading volume. For each firm,主要结果变量是累计异常收益率和异常交易量。对于每家公司，

where $R_{mt}$ is the benchmark return. In the implementation reported here, the benchmark is an equal-weighted portfolio of control software and infrastructure firms, which makes the abnormal-return measure directly interpretable as relative repricing within the adjacent technology-finance ecosystem. Abnormal trading volume is defined as其中 Rm​tR_{mt} 是基准收益率。在此处报告的实施中，基准是对照组软件和基础设施公司的等权重组合，这使得异常收益率指标可以直接解释为相邻技术-金融生态系统内的相对定价调整。异常交易量定义为

where the bar denotes the mean over the estimation window.其中横线表示估计窗口内的平均值。

The design considers narrow and moderate event windows, including $[0,+1]$, $[-1,+1]$, and $[-3,+3]$, with an estimation window of approximately $[-120,-20]$ trading days. Because the February 24, 2026 event generates the strongest market response in the data, the baseline cross-sectional specification focuses on that event:该设计考虑了窄窗口和中等窗口的事件窗口，包括 [0,+1]、[−1,+1] 和 [−3,+3]，估计窗口约为 [−120,−20] 个交易日。由于 2026 年 2 月 24 日的事件在数据中产生了最强的市场反应，基准横截面设定集中在该事件上：

where $Vendor_{i}$ identifies legacy-service providers and $Financial_{i}$ identifies financial-sector firms. This is intentionally a reduced-form specification. The objective is not to claim a complete structural model, but to test whether the sign and magnitude of repricing differ systematically by exposure and business model.其中 V​e​n​d​o​riVendor_{i} 标识遗留服务供应商，F​i​n​a​n​c​i​a​liFinancial_{i} 标识金融部门公司。这特意采用了一种简化形式的设定。其目的不是声称建立一个完整的结构模型，而是测试定价调整的符号和幅度是否因风险敞口和商业模式的不同而存在系统性差异。

The identification logic follows directly from the mechanism above. If investors view agentic AI as a workflow technology rather than only a forecasting technology, then capability disclosures should affect firms differently depending on where they sit in the relevant production chain. Three hypotheses follow.识别逻辑直接遵循上述机制。如果投资者将智能体 AI 视为一种工作流技术，而不仅仅是一种预测技术，那么能力披露应根据公司在相关生产链中所处的位置，对其产生不同的影响。由此得出三个假设。

First, capability disclosures should generate economically meaningful abnormal returns and abnormal trading volume in firms with greater exposure to legacy modernization. Second, the sign of the market reaction should differ by business model: firms that sell legacy-maintenance or labour-intensive modernization services should be more negatively repriced than firms that mainly consume legacy modernization as an input. Third, within the finance vertical, reactions may be stronger when legacy technology debt is more deeply tied to regulated core operations such as payments, claims processing, and core banking.第一，能力披露应在对遗留现代化风险敞口较大的公司中产生具有经济意义的异常收益率和异常交易量。第二，市场反应的符号应因商业模式而异：销售遗留维护或劳动密集型现代化服务的公司，其股价调整应比主要将遗留现代化作为投入品的公司更为负面。第三，在金融垂直领域内，当遗留技术债务与支付、索赔处理和核心银行等受监管的核心业务联系更紧密时，反应可能更强烈。

The design is therefore informative for the broader paper even though it does not observe internal model weights, prompts, or deployment logs. It identifies a more observable margin: how public markets price expected changes in workflow substitution, modernization cost, and technology dependence when agentic capability becomes more credible.因此，尽管该设计并未观察到内部模型权重、提示词或部署日志，但它对整篇论文具有参考价值。它识别出了一个更具可观测性的边际：当智能体能力变得更加可信时，公开市场如何对工作流替代、现代化成本和技术依赖的预期变化进行定价。

To move beyond design and provide a first empirical read, the application was implemented on a 31-firm sample using publicly retrievable price and filing data between June 2025 and March 2026. The sample comprises eight legacy-service vendors, sixteen legacy-heavy financial users, and seven software-or-infrastructure control firms. One candidate firm from the initial seed universe was dropped because the public market-data endpoint did not provide a stable history for the relevant window.为了超越设计层面并提供初步的实证解读，该应用在 31 家公司的样本上进行了实施，使用了 2025 年 6 月至 2026 年 3 月期间可公开获取的价格和申报数据。样本包括 8 家遗留服务供应商、16 家重度依赖遗留系统的金融用户和 7 家软件或基础设施对照公司。初始种子库中的一家候选公司因公开市场数据端点未能提供相关窗口的稳定历史记录而被剔除。

Table 3 reports subgroup means for the three focal dates. The clearest pattern appears around the February 24, 2026 market-validation event. For legacy-service vendors, mean $CAR[-1,+1]$ is approximately $-6.39\%$. For legacy-heavy financial users, the corresponding figure is $-2.61\%$, while the control portfolio is approximately zero by construction because it is also the benchmark. This is the central empirical result of the section: the market reaction is not uniform, and the sharpest negative repricing is concentrated in the group whose revenues are most directly threatened by accelerated modernization.表 3 报告了三个重点日期的子组平均值。最明显的模式出现在 2026 年 2 月 24 日的市场验证事件前后。对于遗留服务供应商，平均 C​A​R​[−1,+1] 约为 −6.39%。对于重度依赖遗留系统的金融用户，相应数字为 −2.61%，而对照组合由于其本身就是基准，其数值按定义约为零。这是本节的核心实证结果：市场反应并不一致，最剧烈的负面定价调整集中在那些收入最直接受到加速现代化威胁的群体中。

The volume response complements the return evidence. Around the same February event, abnormal log volume rises to approximately $0.73$ for service vendors, $0.37$ for financial users, and $0.70$ for controls. This pattern indicates a broad attention shock rather than a narrow single-name move. However, only part of that attention translates into strongly negative prices, again suggesting that the market is distinguishing between business models rather than reacting mechanically to AI news.交易量反应补充了收益率证据。在同一 2 月事件前后，服务供应商的异常对数交易量上升至约 0.73，金融用户为 0.37，对照组为 0.70。这种模式表明这是一次广泛的关注冲击，而非单一公司的局部波动。然而，只有部分关注转化为强烈的负面价格反应，这再次表明市场是在区分不同的商业模式，而不是对 AI 新闻做出机械反应。

The March 5, 2026 labour-market report generates a different configuration. Vendor returns are close to zero on average, while financial users are more negative. This pattern is consistent with a shift from a capability narrative to an organizational adjustment narrative: once the discussion turns from what the tools can theoretically do to how exposure may materialize in labour and workflow terms, firms that must actually reorganize around legacy technology may bear more of the near-term adjustment concern.2026 年 3 月 5 日的劳动力市场报告产生了不同的配置。供应商收益率平均接近于零，而金融用户则更为负面。这种模式与从“能力叙事”向“组织调整叙事”的转变相一致：一旦讨论从工具理论上能做什么转向风险敞口如何在劳动力和工作流方面具体化，那些实际上必须围绕遗留技术进行重组的公司，可能会承担更多的短期调整担忧。

Table 4 sharpens this interpretation. In the February 24 cross section, the coefficient on the legacy-service vendor indicator is $-0.0262$ with a $t$-statistic of approximately $-1.95$, suggesting economically meaningful underperformance for firms whose business models are more exposed to legacy maintenance and modernization revenues. By contrast, the filing-based exposure score is negative but not precisely estimated in this small sample, while the generic financial-sector dummy is close to zero. The empirical implication is important: the market does not appear to price all legacy exposure in the same way. What matters is not simply whether a firm mentions legacy technology, but whether its revenue model is expected to lose or gain from agentic modernization.表 4 加深了这一解读。在 2 月 24 日的横截面中，遗留服务供应商指标的系数为 −0.0262，t 统计量约为 −1.95，表明对于那些商业模式更依赖遗留维护和现代化收入的公司而言，出现了具有经济意义的跑输市场表现。相比之下，基于申报文件的风险敞口评分虽为负值，但在该小样本中估计并不精确，而通用的金融部门虚拟变量则接近于零。这一实证含义很重要：市场似乎并未以相同的方式对所有遗留风险敞口进行定价。重要的不仅是公司是否提到了遗留技术，而是其收入模式是否预期会因智能体现代化而获益或受损。

Taken together, these findings are consistent with the core claim of the AFMM. Capability disclosures about agentic AI appear to trigger market-wide attention, but price effects are heterogeneous and tied to the expected distribution of rents across exposed actors. In that sense, the application supports the broader thesis of the paper: the economic significance of agentic AI depends not only on abstract model capability, but on how specific architectures and workflows map into incumbent business models, modernization costs, and institutional dependencies.总而言之，这些发现与 AFMM 的核心主张一致。关于智能体 AI 的能力披露似乎引发了全市场的关注，但价格效应是异质的，并与暴露参与者之间的预期租金分布相关。从这个意义上讲，该应用支持了本文的更广泛论点：智能体 AI 的经济意义不仅取决于抽象的模型能力，还取决于特定架构和工作流如何映射到在位者的商业模式、现代化成本和机构依赖中。

This application is tightly connected to the theoretical framework developed in the paper. It does not require observing the internal prompts, memory modules, or control logic of every institutional AI agent. Instead, it studies a more observable and financially meaningful margin: how markets update beliefs when publicly disclosed capabilities shift expectations about autonomy, workflow substitution, and infrastructure dependence.该应用与论文中发展的理论框架紧密相连。它不需要观察每个机构 AI 智能体的内部提示词、记忆模块或控制逻辑。相反，它研究了一个更具可观测性和金融意义的边际：当公开披露的能力改变了关于自主权、工作流替代和基础设施依赖的预期时，市场如何更新信念。

In AFMM terms, the application captures a change in perceived autonomy depth ($A$), workflow substitutability, and execution-relevant capability before large-scale institutional diffusion is directly measurable. It therefore functions as a bridge between the architecture discussion in Section 4 and the systemic implications developed in Section 5. If capability news can already reprice adjacent firms through expectations alone, then widespread deployment could plausibly reshape market structure, business-model risk, and concentration dynamics at a larger scale.用 AFMM 的术语来说，该应用捕捉了在能够直接衡量大规模机构扩散之前，感知自主深度 (AA)、工作流可替代性和执行相关能力的变动。因此，它充当了第 4 节架构讨论与第 5 节系统性影响之间的桥梁。如果能力新闻仅通过预期就能重新定价相邻公司，那么广泛的部署完全有可能在更大范围内重塑市场结构、商业模式风险和集中度动态。

The paper’s main claim is not that a single announcement can reveal the full long-run impact of agentic finance. The point is narrower and more defensible. If secondary markets already differentiate across firms when new agentic capabilities become credible, then architecture and workflow design are economically relevant before full deployment is observable. The application therefore complements the conceptual framework by showing one concrete transmission channel from agent design to market outcomes: public capability news changes expected rents, which in turn changes valuation.本文的主要观点并非单一公告就能揭示智能体金融的全部长期影响。其观点更为狭窄且更具辩护性。如果当新的智能体能力变得可信时，二级市场已经在区分不同公司，那么在完全部署可见之前，架构和工作流设计在经济上就已经具有相关性了。因此，该应用通过展示从智能体设计到市场结果的一个具体传导渠道，补充了概念框架：公开的能力新闻改变了预期租金，进而改变了估值。

The results are also consistent with the paper’s emphasis on bounded autonomy rather than a near-term jump to fully autonomous financial markets. The market reaction is concentrated around workflow substitution and modernization cost, not around a wholesale replacement of institutions. This is exactly the pattern one would expect if the first economically significant effect of agentic AI is to reorganize decision pipelines, service models, and operational bottlenecks before it displaces regulated financial authority itself.研究结果也与本文强调的“有限自主权”而非短期内跳跃到完全自主金融市场的观点相一致。市场反应集中在工作流替代和现代化成本上，而不是机构的全面更替。如果智能体 AI 的首个具有经济意义的影响是重组决策管道、服务模型和运营瓶颈，而不是取代受监管的金融权力本身，这正是人们所预期的模式。

This distinction matters for interpretation. A capability shock can affect valuations even when legal, fiduciary, and governance constraints still prevent full delegation. In other words, financial markets can price the consequences of partial automation, co-pilot systems, and controlled execution modules long before they price a world of unconstrained autonomous agents.这种区分对于解读至关重要。即使法律、信托和治理约束仍然阻止完全授权，能力冲击也能影响估值。换句话说，金融市场可以在完全不受约束的自主智能体世界定价之前，就对部分自动化、副驾驶系统和受控执行模块的后果进行定价。

The framework still has clear limits. The AFMM is a reduced-form conceptual model rather than a fully specified equilibrium model of strategic interaction among heterogeneous agents. The empirical component is also intentionally modest: the present exercise is illustrative, based on a small sample, public-data proxies, and a simplified benchmark rather than a final identification design.该框架仍有明确的局限性。AFMM 是一个简化形式的概念模型，而非异质智能体之间战略互动的完全指定均衡模型。实证部分也特意保持谦逊：目前的练习是说明性的，基于小样本、公开数据代理和简化基准，而非最终的识别设计。

The next empirical step is therefore straightforward. Future work can stack multiple capability-shock events, extend the filing-based exposure measure with richer text methods, incorporate intraday data and option-implied measures, and track actual deployment signals across financial institutions. A parallel theoretical step is to formalize how heterogeneous human and AI agents interact under different degrees of autonomy, coupling, and concentration.后续的实证步骤因此非常明确。未来的工作可以堆叠多个能力冲击事件，利用更丰富的文本方法扩展基于申报文件的风险敞口衡量指标，纳入日内数据和期权隐含指标，并跟踪金融机构间的实际部署信号。平行的理论步骤是形式化人类和 AI 智能体如何在不同程度的自主权、耦合度和集中度下进行互动。

An especially promising next stage is to combine compact event studies like the one developed here with AI-assisted stress monitoring and market-narrative clustering from news, filings, and supervisory text. That would preserve the conceptual orientation of the paper while strengthening its empirical reach.一个特别有前景的下一阶段是将此处开发的紧凑型事件研究与 AI 辅助的压力监测，以及来自新闻、申报文件和监管文本的市场叙事聚类相结合。这将保留论文的概念导向，同时增强其实证覆盖范围。

Despite these limitations, the framework now has a clearer empirical anchor. It offers a plausible way to connect architecture, adoption, governance, and repricing within a single research agenda for agentic finance.尽管存在这些局限性，该框架现在拥有了更清晰的实证锚点。它为在智能体金融的单一研究议程中连接架构、采用、治理和定价调整提供了一种可行的方法。