Agentic AI promises sharper market efficiency and liquidity but brings new systemic and regulatory risks; unless governance, monitoring and interpretability tools keep pace, autonomous trading systems could amplify instability.

Main Finding

Agentic AI—systems that set goals, plan, learn, and coordinate autonomously—represents a discontinuous shift for financial markets. It can materially improve execution, liquidity provision, risk monitoring, and adaptive portfolio management, but it also creates novel sources of market instability, non‑stationarity, opacity, and concentration. Realizing net social value requires new evaluation methods, governance mechanisms, and market‑design adjustments tailored to multi‑agent adaptive systems.

Key Points

• Definition and scope

  • Agentic AI is distinct from classical algorithmic trading and generative models: core features are goal-oriented autonomy, continuous reasoning/planning, adaptability (online learning), multi‑agent coordination, and maintenance of long‑term objectives.
  • The survey focuses on systems with genuine agency (autonomy in goal formulation, adaptive learning, strategic coordination), excluding static rule engines and one‑shot predictive models.

• Core technical characteristics

  • Autonomy: reinforcement‑learning formulations for sequential decision making (reward designs often incorporate volatility scaling).
  • Reasoning & planning: ReAct, Reflection-style loops, planning modules that produce multistep strategies rather than single-shot outputs.
  • Memory & knowledge: layered memory (e.g., FINMEM), retrieval‑augmented generation (RAG), prompt‑chaining / sequential knowledge prompting (SKGP / LLMFactor) to maintain context over long horizons.
  • Multi‑agent architectures: specialization of subagents (fact vs. subjectivity paths, trade/execution agents, supervisory agents) and communication/coordination protocols to decompose complex workflows.
  • Tool use and integration: agents use APIs, databases, market venues and other agents as tools; supervisory agents manage delegation and conflict resolution.

• Applications reviewed

  • Trading/execution: coordinated multi‑agent RL for execution strategies and market making.
  • Portfolio management: dynamic rebalancing with regime detection, adaptive risk budgeting.
  • Risk management: automated, continuous stress testing and scenario analysis at scale.
  • Compliance & governance: automated monitoring with audit trails but interpretability challenges remain.

• Benefits vs. risks

  • Potential benefits: improved execution costs, faster detection/responding to regime shifts, scalable stress testing, enhanced liquidity provision in normal conditions.
  • Systemic risks: correlated behavior across agents, feedback loops amplifying volatility, non‑stationarity undermining validation, opacity hindering oversight, concentration of compute/data raising competitive asymmetries.
  • Operational/regulatory challenges: model validation under continuous learning, interpretability of multi‑agent plans, accountability attribution when decisions are decentralized.

• Gaps and research needs identified

  • Evaluation methodologies for adaptive, multi‑agent systems (including out‑of‑sample and counterfactual stress tests).
  • Interpretability techniques suitable for multi‑agent, planning‑based decision chains.
  • Systemic‑risk metrics for correlated adaptive strategies and mechanisms to mitigate fragility.
  • Governance, auditing standards, and market design modifications to preserve stability as autonomy increases.

Data & Methods

• Literature coverage

  • Systematic survey of “agentic AI”, “finance”, and “reinforcement learning” papers indexed on Scopus through Dec 23, 2025, supplemented by relevant preprints (e.g., SSRN, arXiv).
  • Cross‑disciplinary synthesis spanning computer science (agent design, RL, LLM agents), finance (market microstructure, execution), and operations research.

• Analytical approach

  • Taxonomy: formalizes a taxonomy of agency dimensions (autonomy, reasoning/planning, adaptability, multi‑agent coordination, long‑term goals).
  • Design‑pattern synthesis: catalogs architectural patterns used in agentic financial systems (Prompting, Tool Use, Memory, Planning/Reasoning, Self‑Reflection, Multi‑Agent Collaboration).
  • Comparative analysis: contrasts agentic AI with classical algorithmic trading and generative AI along capability axes and implications.
  • Systemic assessment: maps agentic properties to market‑level outcomes (liquidity, price discovery, resilience, concentration) and regulatory frameworks.

• Empirical content

  • The paper is a survey/synthesis; it does not present new proprietary market microdata or original econometric estimation. It references representative empirical and simulation studies (e.g., RL in trading, multi‑agent market simulations) and several architectural case studies (Zhu et al. 2023; Wang et al. 2024; Yu et al. 2023; Pippas et al. 2024; Zhang et al. 2020).

Implications for AI Economics

• Market microstructure and dynamics

  • Endogenous non‑stationarity: agentic agents that learn online change the data‑generating process—standard inference and backtesting methods become unreliable.
  • Correlated strategies and fragility: multi‑agent coordination and shared training/data sources can induce herding, amplifying liquidity dry‑ups and extreme price moves (higher tail risk).
  • Price discovery & liquidity: agentic agents may improve efficiency in normal regimes but exacerbate illiquidity in stress due to synchronized de‑risking or withdrawal of liquidity provision.

• Competition and concentration

  • Barriers to entry: scale advantages from compute, data, and engineering create winner‑takes‑most dynamics; incumbents with superior agentic stacks can dominate liquidity provision and informational edges.
  • Market power effects: concentrated agentic deployment can alter spreads, trading costs, and market access—calling for scrutiny on competitive impacts.

• Regulation, governance, and policy design

  • New validation and oversight requirements: regulators should require continuous validation frameworks that account for online learning, including mandatory sandbox testing, continuous monitoring, and provenance/audit logs for agent decisions.
  • Explainability & accountability: impose standards for interpretable reporting of agent objectives, decision chains, and fallback/override mechanisms (human‑in‑the‑loop thresholds).
  • Systemic safeguards: consider market‑level interventions—agent registries, simulations of multi‑agent stress scenarios, pre‑deployment “red teaming”, adaptive circuit breakers, and coordination protocols across venues to limit synchronized exits.
  • Disclosure and transparency: firms should disclose agentic agent usage, risk budgets, learning rates, and coordination links where feasible to enable systemic oversight.

• Research agenda for AI economics

  • Theoretical models: incorporate agentic, adaptive agents into equilibrium/market‑design models to analyze endogenous liquidity, volatility, and welfare effects.
  • Simulation and stress testing: develop multi‑agent market simulators with realistic microstructure to study counterfactuals, policy interventions, and emergent behaviors.
  • Econometrics for non‑stationary markets: create identification strategies and validation protocols robust to policy‑driven and agent‑driven regime shifts.
  • Governance economics: design incentive‑compatible regulation and mechanism design that align agent objectives with social welfare (e.g., taxes/subsidies on destabilizing behaviors, certification regimes).
  • Measurement of concentration/externalities: quantify how compute/data concentration translates into market power and systemic risk.

Actionable takeaways for researchers and policymakers - Researchers: prioritize multi‑agent simulation frameworks, validation methods for online learning, and economic theory that models adaptive strategic agents. - Policymakers/regulators: implement continuous validation, require provenance/audit trails for agentic decisions, mandate stress tests that simulate multi‑agent coordination failures, and monitor concentration metrics in compute/data infrastructure. - Practitioners: adopt layered memory and supervisory architectures, conservative online‑learning protocols (guardrails for catastrophic drift), comprehensive logging for auditability, and cross‑firm coordination mechanisms to avoid synchronized destabilizing actions.

Summary conclusion Agentic AI offers substantial operational upside for financial markets but fundamentally alters the dynamics that underlie market stability and competition. Managing the transition demands coordinated advances in technical evaluation, economic theory, and regulatory design tailored to adaptive, multi‑agent systems.