Large language model (LLM) trading agents can ingest market data, news, and executable analyses, yet their behavior is usually governed by static hand‑crafted tool‑use policies fixed before deployment. Such rigidity limits the agents' ability to adapt evidence gathering, tool invocation, signal verification, and risk management as market regimes shift.\ \ EvolveTrade treats the system prompt of a tool‑using trading agent as a text‑parameterized policy. After each update interval, a Policy Agent revises this policy using accumulated decision traces and realized portfolio feedback, while keeping the backbone LLM unchanged. The revised policy then drives the next batch of trading decisions, allowing the agent to progressively refine its information‑acquisition and portfolio‑construction procedures over time.\ \ Experiments across several market regimes and two different LLM backbones show that EvolveTrade consistently improves Sharpe Ratio and Cumulative Return compared with fixed‑policy baselines, achieving gains in most evaluated settings. Behavioral analyses reveal that self‑evolved policies increase code‑mediated analysis, activate regime‑relevant computations, and case‑level policy‑to‑return attributions trace how policy‑induced allocation shifts translate into realized return differences.\ \ These findings suggest that adapting the reusable tool‑use procedure is a crucial direction for building more robust LLM trading agents.\ \ Review