Agentic trading systems extend financial automation beyond isolated prediction by coordinating evidence acquisition, research, belief formation, portfolio construction, risk control, execution, memory, and post-trade learning within a governed decision loop. This paper develops a formal architecture and control framework for such systems. It separates evidence, cognition, control, execution, and assurance into distinct planes; formalizes a bounded market agent with revocable authority; proves that the execution topology is non-bypassable and that the governed action is stable against bounded adversarial perturbation; and derives uncertainty-aware portfolio, safe-adaptation, intent-to-order, and operational-state mechanisms. The central claim is that autonomy is not a binary model property but a state-dependent capital permission that must contract when calibration, regime familiarity, data quality, drawdown, security, or operational reliability deteriorate. The paper also synthesizes the emerging empirical record and explains why fluent reasoning or headline profitability does not constitute deployment evidence. A computational experiment implements specialist agents, regime detection, adaptive authority, conformal calibration, volatility and CVaR constraints, order compilation, transaction costs, tamper-evident audit, adversarial fault injection, governance ablations, and deterministic replay. The experiment is illustrative rather than an alpha claim: across two hundred independent regime paths the governed system reduces volatility, tail loss, and drawdown relative to equal weight in almost every path, while on the reference path strong learned baselines (mean-variance and Black--Litterman), run under identical costs and controls, beat it on Sharpe ratio and return. The contribution of governance is therefore an auditable risk envelope rather than excess return, and a certainty-equivalent analysis identifies the risk-aversion level above which that trade is worthwhile. The resulting framework treats agentic trading systems as governed artificial institutions whose capital authority is revocable, measurable, and attributable.
Alonso et al. (Mon,) studied this question.