The Commonplace
Home Dashboard Papers Evidence Digests 🎲
← Papers

AI will reshape market dynamics through deployment and governance choices rather than raw model power; supervised 'co-pilot' architectures are the likely near-term norm, lowering some immediate risks but creating new coupling and concentration vulnerabilities.

AI Agents in Financial Markets: Architecture, Applications, and Systemic Implications
Hui Gong · March 14, 2026 · ArXiv.org
openalex theoretical low evidence 7/10 relevance Source PDF
The systemic impact of AI agents in financial markets hinges more on how agent architectures are distributed, coupled, and governed across institutions than on model intelligence alone, with bounded autonomy (supervised co-pilots and constrained execution) the most plausible near-term equilibrium.

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 develops an illustrative empirical application based on event studies of AI-agent capability disclosures and heterogeneous market repricing. The central argument is 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. 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.

Summary

Main Finding

The paper develops an integrated framework for “agentic finance” and argues that the systemic effects of AI in financial markets will depend less on raw model intelligence than on how agentic systems are architected, coupled, distributed, and governed. It introduces a four‑layer agent architecture and an Agentic Financial Market Model (AFMM) that maps agent design choices (e.g., autonomy depth, model heterogeneity, execution coupling, infrastructure concentration, supervisory observability) to market outcomes (efficiency, liquidity resilience, volatility, herding, concentration risk, supervisory capacity). In the near term the most likely equilibrium is “bounded autonomy” (supervised co‑pilots and constrained execution); however, particular design and infrastructure patterns could amplify systemic vulnerabilities (correlated behaviour, fragility, third‑party concentration).

Key Points

  • Definitions & framing
    • Agentic finance: markets where semi‑ or fully‑autonomous AI systems materially participate in information processing, decision support, monitoring or execution.
    • Shift from model‑centric automation (prediction) to workflow‑centric automation (perception → reasoning → strategy → execution).
  • Four‑layer agent architecture
  • Data perception (heterogeneous structured/unstructured inputs, retrieval, memory)
  • Reasoning engines (LLMs, planners, constraint handling)
  • Strategy generation (portfolio/trading/risk policies, multi‑step plans)
  • Execution with control (order routing, limits, human supervision, execution coupling)
  • Agentic Financial Market Model (AFMM)
    • Links micro design parameters (autonomy depth, model heterogeneity, execution coupling, infrastructure concentration, supervisory observability) to macro outcomes (price efficiency, liquidity provision and resilience, volatility dynamics, herding/correlation, market concentration, supervisory load).
    • Proposes mechanism‑based propositions (e.g., high execution coupling + low heterogeneity → amplified volatility/herding; high infrastructure concentration → third‑party systemic risk).
  • Three generations of financial AI
    • Algorithmic finance (rule automation / execution);
    • Machine‑learning finance (prediction tasks);
    • Agentic finance (workflow automation, planning, tool use).
  • Application domains
    • Trading & execution, portfolio construction, risk monitoring & compliance, DeFi analytics and automation.
  • Likely near‑term equilibrium
    • Bounded autonomy: AI as supervised co‑pilots, constrained execution modules, monitoring aids.
  • Risks and benefits
    • Benefits: improved monitoring, faster information aggregation, richer analytics, compliance automation.
    • Risks: correlated machine‑mediated behaviour, liquidity fragility, amplified intraday volatility, vendor concentration, supervisory blind spots.
  • Research & policy agenda
    • Need for empirical measures of adoption, heterogeneity, coupling and concentration; simulation and stress‑testing; governance rules for observability and execution limits.

Data & Methods

  • Methodological stance
    • Primarily conceptual and theory‑building: synthesises six literature streams (agent‑based finance, market microstructure, financial LLMs, agent architectures, AI in trading, regulation/systemic risk) and derives mechanism‑based propositions linking agent design to market outcomes.
  • Proposed empirical toolkit (outlined, not implemented in paper)
    • Observable market data: liquidity measures, intraday volatility, order‑book dynamics, trade timestamps, market microstructure metrics.
    • Textual & model signals: public filings, news, social media, model outputs/summaries, adoption signals (job postings, vendor contracts, software telemetry where available).
    • Identification approaches: cross‑sectional and panel comparisons across firms/venues with different agent architecture proxies; event studies around AI deployment announcements; agent‑level and market‑level LLM/agent simulations (e.g., LLM agent market simulations like ASFM/TwinMarket) for counterfactuals.
    • Metrics to operationalise AFMM parameters: proxies for autonomy depth (degree of executed trades without human sign‑off), model heterogeneity (variety of data sources and model families used), execution coupling (latency and direct market access measures), infrastructure concentration (market share of third‑party model/hosting vendors), supervisory observability (audit logs, human‑in‑loop ratios where available).
  • Illustrative/empirical agenda
    • The paper outlines an illustrative empirical application showing how to combine market microstructure data with textual indicators and public adoption proxies to test AFMM propositions (e.g., link adoption intensity + execution coupling to changes in liquidity resilience and correlated trading).

Implications for AI Economics

  • New micro → macro causal channels
    • AI economics must incorporate agent architecture, coupling, and governance variables as causal determinants of macro market outcomes, not just model accuracy.
  • Measurement priorities
    • Develop firm‑ and market‑level measures of autonomy, heterogeneity, coupling, and infrastructure concentration; collect observability/telemetry standards to enable evaluation.
  • Theory & modelling
    • Extend heterogeneous‑agent and market‑microstructure models to include agentic decision pipelines (perception → reasoning → action) and endogenous information flows from agent coordination, memory and tool use.
    • Use LLM/agent simulations and controlled field experiments to study propagation, amplification, and feedback effects.
  • Policy & regulation
    • Regulatory focus should shift from model accuracy alone to: limits on delegated execution, requirements for observability/auditability, third‑party concentration monitoring, stress‑testing agentic behaviours, and guidelines for human‑in‑the‑loop controls.
    • Macroprudential tools may be needed to address vendor concentration and correlated agent behaviour.
  • Research agenda for economists
    • Empirically identify how agent design parameters affect liquidity, volatility, and systemic risk; quantify tradeoffs between efficiency gains and stability costs.
    • Study market equilibria under bounded vs. greater autonomy and potential tipping points where autonomy, coupling and concentration produce nonlinear systemic effects.
    • Evaluate policy interventions (transparency mandates, execution caps, third‑party resilience requirements) through empirical analysis and simulation.

Short summary: the paper provides a modular, testable framework that shifts attention in AI‑finance research from isolated model performance to the institutional and architectural features of agent deployment—features that are pivotal for market‑level outcomes and for the design of effective regulation.

Assessment

Paper Typetheoretical Evidence Strengthlow — The core contribution is theoretical and model-based; the empirical component is presented as illustrative event studies of capability disclosures which can show contemporaneous market reactions but do not establish robust causal mechanisms or long-run economic effects, and appear limited in scope and depth. Methods Rigormedium — The paper provides a clear, structured theoretical architecture and a transparent stylised agent-based model linking micro design choices to macro outcomes, which is methodologically sound for a conceptual contribution; however, the empirical work is limited to illustrative event studies without stronger identification, robustness checks, or detailed microdata on actual trading behavior and governance practices. SampleTheoretical development and a stylised agent-based model; empirical illustrations use event-study-style analyses of market price reactions to firm disclosures about AI-agent capabilities, comparing heterogeneous repricing across firms/sectors (stock market data around announcement windows). Themesgovernance adoption IdentificationPrimarily conceptual/analytical: develops a four-layer architecture and a stylised agent-based model (Agentic Financial Market Model) to map agent design parameters to market outcomes; includes an illustrative empirical event-study approach that examines market repricing around firm disclosures of AI-agent capabilities to suggest heterogeneous effects, relying on standard event-study assumptions (announcement exogeneity and efficient markets) rather than a tight causal identification strategy. GeneralizabilityStylised agent-based model abstracts from many market microstructure details and institutional specifics, Event-study evidence is limited to announcement-period price reactions and may not reflect realized trading behavior or long-run outcomes, Findings may be concentrated in public equity markets and not directly applicable to OTC, fixed income, or private markets, Context- and jurisdiction-specific governance, regulation, and infrastructure differences limit transferability across countries, Heterogeneity in firm size, access to infrastructure, and algorithmic sophistication may constrain generalization to all market participants

Claims (6)

ClaimDirectionConfidenceOutcomeDetails
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. Adoption Rate positive high shift in type of financial automation (from isolated prediction to integrated decision systems)
0.06
Financial AI agents can be described by a four-layer architecture covering data perception, reasoning engines, strategy generation, and execution with control. Other mixed high architectural decomposition of financial AI agents
0.02
The Agentic Financial Market Model (AFMM), a stylised agent-based representation, links agent design parameters (autonomy depth, heterogeneity, execution coupling, infrastructure concentration, supervisory observability) to market-level outcomes including efficiency, liquidity resilience, volatility, and systemic risk. Market Structure mixed high market-level outcomes (efficiency, liquidity resilience, volatility, systemic risk) as functions of agent design parameters
0.02
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. Market Structure mixed high systemic implications / market-level risk and stability as a function of architecture distribution, coupling, and governance
0.12
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. Adoption Rate neutral high expected equilibrium mode of AI agent autonomy in finance (bounded autonomy / supervised co-pilot)
0.02
The paper develops an illustrative empirical application based on event studies of AI-agent capability disclosures and heterogeneous market repricing. Market Structure mixed medium market repricing heterogeneity following AI-agent capability disclosure events
0.04

Notes