Liz Agent Research: AI Best Practices for Investment Advisors

Overview

Comprehensive research report on AI agent best practices for the Liz trading assistant. Full report at ~/clawd/projects/hedge/docs/LIZ_AGENT_RESEARCH.md (716 lines, 35KB).

Key Findings

Architecture

• Liz uses the correct ReAct (Reason+Act+Observe) pattern — tool-use agentic loop
• Best practice: prefer atomic tools for data retrieval, composite tools for common workflows
• Multi-agent financial systems (e.g., MarketSenseAI) use specialized sub-agents: macro, sector, fundamental, technical, sentiment → synthesis

Top 10 Improvements (Prioritized)

1. Extended thinking (LOW effort, VERY HIGH impact) — enable budget_tokens: 8000-16000 for complex analysis
2. Reasoning scaffold in system prompt — add explicit step-by-step analysis instructions
3. Cross-session memory — Liz forgets everything between sessions; add Supabase persistence
4. Optimized tool responses — add summary field to every tool response for LLM consumption
5. Increase max_tokens (2000 → 8000) + add streaming
6. Bias prevention — add explicit uncertainty/bear-case requirements to system prompt
7. Proactive alerts — scheduled signal scanning, regime change alerts (HIGH effort)
8. Prompt caching — 40-60% cost reduction on static system prompt + tools
9. Few-shot examples — 2-3 example exchanges in system prompt
10. Composite tool bundles — get_trade_setup(), get_morning_briefing(), get_401k_status()

Data Sources Research (JPMorgan ICAIF 2024)

• Fundamental data (earnings, revenue, cash flow) = most important
• Sentiment SCORES > full news text (equal performance, fewer tokens, less bias)
• Omitting news entirely often improves performance by reducing recency bias

Financial Data Formatting for LLMs (Daloopa 2025)

• LLMs lose table structure when data is flattened to text
• Use hierarchical JSON with preserved relationships
• Include human-readable interpretations alongside raw numbers
• Always include summary field with 1-2 sentence interpretation

Prompt Engineering for Finance

• Chain-of-Thought prompting dramatically improves financial analysis accuracy
• Analogy-Driven CoT (AD-FCoT) grounds reasoning in historical precedents
• FinCoT (2025) injects domain-specific reasoning blueprints at each step
• Explicit bias prevention: always require bear case, uncertainty levels, invalidation conditions

Current Liz Gaps

• No cross-session memory (biggest UX gap)
• max_tokens = 2000 too low for complex analysis
• No reasoning scaffold in system prompt
• No explicit bias prevention instructions
• Tool responses not optimized for LLM consumption (raw API JSON)
• No proactive alerting
• No prompt caching (paying full token cost per call)

Sources

25+ sources including: Anthropic building-effective-agents, OpenAI agent SDK, JPMorgan ICAIF 2024, MarketSenseAI 2025, FinCoT 2025, AD-FCoT 2025, Daloopa financial data guide, AWS financial AI patterns, RAGflow 2025 review, ArXiv LLM investment bias paper, McKinsey agentic evaluation.