Multi-Agent LLM Trading System: Fine-Grained Task Decomposition Beats Coarse Instructions
Multi-agent LLM trading systems are proliferating, but most give models vague role descriptions for decision-making. This paper demonstrates coarse-grained role-playing is insufficient—investment analysis must be explicitly decomposed into fine-grained sub-tasks: macro analysis, financial statements, technical analysis, news sentiment, and integrated decisions. Each handled by a specialized Agent with defined I/O.
In leakage-controlled backtesting on Japanese stocks, fine-grained decomposition significantly outperforms coarse-grained designs. The key driver isn't individual analysis quality but alignment between outputs and decision preferences. For agentic AI in AI trading and other specialized domains, explicit task decomposition beats role-playing.
Many LLM trading systems assign Agents vague roles like "analyst" and "fund manager." This paper argues: **coarse-grained role-playing isn't enough — you need fine-grained task decomposition**.
System Design
The framework decomposes investment analysis into explicit sub-tasks: macro analysis, financial statement analysis, technical analysis, news sentiment, and integrated decision-making. Each is handled by a specialized Agent with defined I/O formats.
Experimental Setup
Tested on Japanese stock market data (TSE) with price data, financial statements, Nikkei/Reuters news, and macro indicators. Strict leakage-controlled backtesting with temporal data splits.
Key Findings
1. Fine-grained decomposition **significantly outperforms** coarse-grained role-playing on risk-adjusted returns
2. The critical performance driver is **alignment between analytical outputs and decision preferences** — not analysis quality alone, but whether results are effectively utilized by the decision module
3. Portfolio optimization exploiting low correlation with the index yields superior Sharpe ratios
Implications
For Agent applications in specialized domains, "you are an analyst" prompts are insufficient. Professional workflows must be explicitly modeled as structured task chains.
Agentic AI in Finance
One of 2026's biggest agentic AI debates: can AI Agents truly handle complex professional work? This paper offers a conditional "yes"—the key is task architecture design. AI trading is one of agentic AI's earliest production domains, thanks to clear feedback signals (P&L) and quantifiable evaluation metrics (Sharpe ratio). The fine-grained task decomposition approach here can be extended to multi-agent system design in other specialized domains.
In-Depth Analysis and Industry Outlook
From a broader perspective, this development reflects the accelerating trend of AI technology transitioning from laboratories to industrial applications. Industry analysts widely agree that 2026 will be a pivotal year for AI commercialization. On the technical front, large model inference efficiency continues to improve while deployment costs decline, enabling more SMEs to access advanced AI capabilities. On the market front, enterprise expectations for AI investment returns are shifting from long-term strategic value to short-term quantifiable gains.
However, the rapid proliferation of AI also brings new challenges: increasing complexity of data privacy protection, growing demands for AI decision transparency, and difficulties in cross-border AI governance coordination. Regulatory authorities across multiple countries are closely monitoring these developments, attempting to balance innovation promotion with risk prevention. For investors, identifying AI companies with truly sustainable competitive advantages has become increasingly critical as the market transitions from hype to value validation.