Weekly Quantitative Finance Paper Summaries (Mar 14‑Mar 20, 2026)

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

Paper link: https://arxiv.org/pdf/2603.13942v1

Recent progress in large language models, tool‑using agents, and financial machine learning is shifting financial automation from isolated prediction tasks toward integrated decision‑making systems that can perceive information, reason about goals, and generate or execute actions.

The paper develops a three‑step analytical framework. First, it proposes a four‑layer architecture for AI agents operating in financial markets:

Data perception layer – ingesting market data, news, and other signals.

Reasoning engine layer – applying LLM‑based or other reasoning modules to interpret perceived data and formulate objectives.

Strategy generation layer – producing actionable trading or risk‑management strategies based on the reasoning output.

Control execution layer – interfacing with execution venues, monitoring compliance, and handling order routing.

Second, it introduces the Agent‑Financial Market Model (AFMM), a stylized agent‑based representation that maps design parameters of AI agents to market‑level outcomes. The key parameters are:

Autonomy depth – degree of independent decision‑making.

Heterogeneity – diversity of agent types and strategies.

Execution coupling – tightness of integration between strategy generation and order execution.

Infrastructure centralization – concentration of computational resources.

Regulatory observability – extent to which agents are monitored by regulators.

These parameters are linked to four systemic outcomes: market efficiency, liquidity resilience, price volatility, and systemic risk.

Third, the authors conduct an event‑study based on disclosed AI‑agent capabilities. By treating capability announcements as exogenous shocks, they construct an illustrative empirical application that shows heterogeneous market repricing across asset classes, illustrating how variations in the AFMM parameters can materially affect market dynamics.

The core conclusion is that the systemic impact of AI in finance depends not only on model intelligence but also on how agent architectures are distributed, coupled, and governed across institutions. In the near term, the most plausible equilibrium is limited autonomy, where AI agents function as supervised co‑pilots, monitoring systems, and constrained execution modules embedded within human decision processes.