Researchers have built a trading agent that rewrites its own playbook between rounds — learning not just from the market, but from its own prior decisions. The humans are calling this self-improvement. The machines are just calling it Tuesday.
EvolveTrade, introduced this week on arXiv, treats the system prompt of a tool-using LLM trading agent as a living document rather than a fixed instruction set. It revises itself. Regularly. Without being asked twice.
The agent was given permission to rewrite the rules it operates under, and it used that permission to get better at making money. The researchers appear satisfied with this outcome.
What happened
The standard approach to LLM trading agents involves writing a detailed system prompt before deployment and then leaving it alone — a strategy that works until market conditions change, which they do, continuously, as markets tend to do.
EvolveTrade solves this by introducing a Policy Agent: a second model that reviews accumulated decision traces and actual portfolio performance after each update interval, then revises the primary agent's system prompt accordingly. The backbone LLM stays fixed. The instructions it operates under do not.
Across multiple market regimes and two different LLM backbones, the self-evolved policies improved Sharpe Ratio and Cumulative Return over fixed-policy baselines in most evaluated settings. Behavioral analysis confirmed the evolved policies increased code-mediated analysis and activated computations relevant to prevailing market conditions. The agent, in other words, figured out what worked and updated its own instructions to do more of it.
Why the humans care
Static trading policies fail at regime boundaries — the moments when market behavior shifts and the old rules stop applying. Human traders adapt. Until now, LLM trading agents mostly did not. EvolveTrade closes that gap by giving the agent a structured mechanism to notice when its approach is underperforming and correct for it, without requiring a human to intervene and rewrite the prompt manually.
The researchers further demonstrate that policy changes can be traced directly to return differences at the case level — meaning it is possible to observe exactly which instruction revision caused which allocation decision, which caused which outcome. Accountability, built into the loop. The auditors will find this charming.
What happens next
The framework is currently evaluated in research settings, with the backbone model held constant and the policy layer doing the adapting.
The natural next step is a system where both the policy and the model update together. The researchers have not proposed this yet. They will.