finance-agent-architectureslisted

How the mainstream finance agent systems are built, and how to stage a research-to-execution pipeline whose gates are code. TRIGGER - build a multi-agent trading system; TradingAgents architecture; the ai-hedge-fund repo; RD-Agent for quant; Vibe-Trading; FinRobot vs FinGPT; FinMem layered memory; a LangGraph, CrewAI, AutoGen or Claude Agent SDK pipeline for stock research; analyst, researcher, trader and risk-manager agents, bull-bear debate; an agent that reads 10-Ks and trades; "how should the pipeline be staged", where the LLM sits, human-in-the-loop gates, prompt injection through scraped filings, agent reproducibility; 交易 agent 架构, 多智能体 pipeline. SKIP for whether any of it makes money or whether you should build one at all (llm-finance-agents), choosing an MCP server (finance-mcp-servers), RL agents (rl-and-ml-trading), and order safety at the broker (broker-execution-apis).
howard-lynn-ye/fin-skills · ★ 1 · AI & Automation · score 77

Install: claude install-skill howard-lynn-ye/fin-skills

# Finance agent architectures and pipelines Every mainstream system has the same shape: **roles copied from a trading firm's org chart, wired by an orchestration graph, fed by tool calls, with a memory/reflection loop.** They differ on two axes that matter more than the role names — *where the LLM sits* (research time, decision time, or execution time) and *what is code versus prompt*. None of them ships the hard gates that decide whether the output may be reported or traded. Those come from this repo. This skill is the architecture question. Whether any of it makes money is `../llm-finance-agents/SKILL.md`; read its §1 first, because the honest answer shapes the design: the LLM belongs at research time, behind gates, until evidence says otherwise. ## 1. The mainstream systems (✅ verified 2026-09-08 — details in `references/systems.md`) | System | ★ · pushed · licence | Orchestration | Roles | Memory | LLM sits at | Lacks | |---|---|---|---|---|---|---| | **TradingAgents** `TauricResearch` | 103,315 · 09-07 · Apache-2.0 | LangGraph `StateGraph` | 4 analysts → bull/bear debate → research mgr → trader → 3-way risk debate → PM | markdown decision log + reflection on realised alpha | **decision** (5-tier rating is the output) | backtester, costs, execution, kill switch | | **ai-hedge-fund** `virattt` (v2.2.0, PyPI `aihf`) | 63,289 · 09-03 · MIT | plain `run_cycle` pipeline, LangChain provider clients | investor personas → blend → hard risk clamps → broker | prompt cache (exac