Agents Are Not Algorithms: The Tradeoffs of Decision-Time Reasoning in AI Trading

Abstract

Agentic AI systems built on large language models reason about each decision in real time rather than executing a pre-specified policy. We study how this property affects trading performance in a real-time market simulator where the agent is tasked with tender selection and execution. We experimentally vary reasoning intensity and market speed for agents based on frontier models. Additional reasoning improves decision quality on some margins but consumes time while the market evolves, generating a tradeoff that depends on market speed. A reasoning dividend appears in tender selection and conditional execution; a deliberation tax appears in stuck inventory and attempts to accept expired tenders. The tradeoff also depends on what is inside the agent's real-time decision loop versus what is compiled into the surrounding system, with a deterministic algorithm as the fully compiled limit case. The agentic architecture becomes competitive with this fully compiled benchmark when the agent is reserved for tender judgment and supported by pre-computed calculations and compiled unwind execution. Agents are not algorithms, and reasoning is not free.

Keywords: Agentic trading, artificial intelligence, large language models, algorithmic trading

JEL Classification: G10, G11, G12