Governing Agentic Finance: A Framework for EU AI Act Compliance in Autonomous Algorithmic Trading

Large language models (LLMs) and multi-agent reinforcement learning (MARL) are enabling systems that interpret information, formulate strategies, execute trades, and manage risk. These systems create challenges involving transparency, traceability, human oversight, operational resilience, and destabilizing interactions among trading agents. This paper proposes an Agentic Compliance Architecture that translates legal obligations and governance principles into controls for LLM-enabled, MARL-based trading systems. The paper does not assume that trading agents are categorically classified as high-risk systems under the European Union Artificial Intelligence Act (EU AI Act). Applicability depends on intended purpose and deployment context. For systems falling within the high-risk framework, the architecture maps requirements under Articles 9, 12, 13, 14, and 15 to controls; where provisions are not directly binding, principles are treated as normative governance-by-design requirements. The analysis also considers DORA requirements applicable to financial entities, particularly ICT risk management, operational resilience, monitoring, and resilience testing. The empirical design uses a simulated electronic market with 20 agents across five strategies, five equity instruments, and five market regimes. Thirty replications for each condition-regime combination yield 300 simulation runs. For each strategy type, two agents operate without the compliance layer and two agents use the trading process with pre-trade risk limits, decision logging, explanation requirements, human-intervention thresholds, and compliance monitoring. Performance is evaluated using Sharpe ratio, maximum drawdown, profitability, volatility, transaction costs, and execution latency. Governance is operationalized through explanation quality, traceability coverage, intervention responsiveness, and compliance-violation rates. Systemic-risk dynamics are proxied using volatility stress, liquidity deterioration, price dislocation, cross-agent return correlation, drawdown contagion, and cascading-loss events, forming a Systemic Risk Index. The framework tests whether governance controls reduce compliance violations and agent destabilization while preserving performance and execution speed. The study provides a design for evaluating trade-offs between trading performance and responsible AI governance without completed findings.