Research
A new UCL framework for modeling AI trading agents finds that systemic risk does not rise smoothly as autonomy increases, it rises non-linearly once autonomy and infrastructure concentration outpace how well supervisors can actually observe what the agents are doing
Hui Gong's Agentic Financial Market Model links five design parameters, autonomy depth, model heterogeneity, execution coupling, infrastructure concentration, and supervisory observability, to market-level outcomes, and concludes the least risky near-term equilibrium is "bounded autonomy": AI agents as supervised co-pilots rather than fully independent market actors.
Most of the public debate about AI trading risk treats autonomy as a dial: more autonomous agents, more risk, in roughly linear proportion. Gong's model says that framing is wrong in a way that matters for how regulators and firms should actually calibrate deployment. Risk does not creep up steadily as agents get more independent, it stays roughly manageable for a while and then rises sharply once autonomy depth and infrastructure concentration outpace the market's ability to actually see what the agents are collectively doing. That threshold effect, not the average level of autonomy, is the thing worth designing controls around.
The model's five parameters are doing real conceptual work, not just providing academic texture. Execution coupling and infrastructure concentration, essentially how many agents share the same execution pathways and cloud or brokerage infrastructure, are flagged as the pair that most directly amplifies synchronized responses: when many agents share infrastructure and react to the same signals through similar pathways, a stress event does not stay isolated, it propagates through the shared plumbing. Model heterogeneity cuts the other way, more diverse model architectures and strategies across agents actually improve price discovery rather than degrading it, which argues against the instinct to standardize AI trading tools for the sake of predictability. Standardization concentrates the failure mode it is meant to prevent.
The event-study component adds an empirical anchor that pure theory papers usually lack: using public disclosures about AI's capacity for legacy code modernization, Gong finds market repricing already differs measurably across firms based on their exposure to AI-driven workflow substitution. That is evidence the market is already pricing agentic AI's second-order effects, not just the headline productivity story, well ahead of most regulatory frameworks catching up to the same distinction.
The five-part governance recommendation, bounded autonomy, traceability, diversity-by-design, embedded intervention capacity, and supervisory co-evolution, reads as a direct answer to FINRA's 2026 oversight report land on the same week: regulators are separately arriving at the conclusion that supervision has to evolve alongside the technology rather than being bolted on after deployment. Where the paper is more useful than the regulatory guidance is in giving that intuition actual structure. "Watch for synchronized behavior across shared infrastructure" is a specific, testable thing a risk team can build a monitor for. "Supervise AI appropriately" is not.
Read the original: arXiv (UCL Institute of Finance & Technology) - A new UCL framework for modeling AI trading agents finds that systemic risk does not rise smoothly as autonomy increases, it rises non-linearly once autonomy and infrastructure concentration outpace how well supervisors can actually observe what the agents are doing. Commentary is the independent editorial view of Share Trading; the original article is credited to its publisher.