TradingAgents-CN: A Deep Migration Plan from LangGraph to Agno

This is a structured English summary of a long Chinese technical planning document.

1. Background

TradingAgents-CN is a multi-agent financial trading decision framework:

• Stack: LangGraph 0.4.8 + LangChain + FastAPI (+ MongoDB/motor, Redis)
• Agents: 11 core agents — 4 analysts (Market, SocialMedia, News, Fundamentals), 2 researchers (Bull, Bear), 1 Trader, 4 risk roles (Risky/Safe/Neutral debators, RiskManager)
• Data: multi-source integration for China A-shares, HK stocks, US stocks (akshare, tushare, yfinance, finnhub, eodhd, baostock)
• LLM providers: OpenAI, Anthropic, DeepSeek, Google, DashScope (Ali Bailian), etc., via a factory pattern
• Deployment: full Web app + CLI + API service

Current graph workflow uses StateGraph with a shared AgentState (extending MessagesState) and typed sub-states like InvestDebateState (bull/bear history, judge decision).

2. Migration Motivation

1. Performance: Agno claims agent creation ~10,000x faster and ~1/50 the memory usage vs LangGraph 2. Architecture: decentralized execution engine, zero-copy data pipeline 3. Developer efficiency: declarative, modular agent definitions 4. Multimodal support and alignment with agent-framework trends

Goals: 1000–10000x agent creation speedup, 90–95% memory reduction, full functional parity, backward-compatible APIs.

3. LangGraph vs Agno (author's assessment)

| Dimension | LangGraph | Agno | |---|---|---| | Agent creation speed | baseline | claimed 10,000x | | Memory usage | baseline | claimed 1/50 | | Learning curve | moderate | simpler | | Ecosystem | mature | emerging | | Feature completeness | complete | unverified |

Notably, the author flags Agno's maturity, API compatibility, and documentation as high risks; the Agno architecture samples in the document are explicitly speculative.

4. Recommended Strategy and Target Architecture

Three strategies were compared: progressive migration (recommended), full rewrite, and dual-track running. The recommended target architecture includes:

• TradingAgentsWorkflow(Workflow) with an AnalysisStage enum; analysts run in parallel, researchers sequentially, risk debators in parallel, trader last
• Declarative TradingState, DebateState, RiskDebateState classes with optional report fields and debate stages
• Decorator-based @Tool definitions for stock data, technical indicators, news, valuation
• An AgnoLLMFactory producing provider-specific LLM instances from environment API keys

5. Phased Plan (~12–18 weeks, 6 stages)

1. Base infrastructure (2–3 wks): Agno setup, state system, LLM adapter layer, minimal agent template 2. Core agents (3–4 wks): migrate 4 analysts + 2 researchers, tool system, prompt templates 3. Workflow orchestration (2–3 wks): workflow engine, agent messaging, parallel/async coordination, state propagation and rollback 4. Advanced features (2–3 wks): memory system (FinancialSituationMemory + vector DB), risk management, agent pooling, monitoring 5. Integration & testing (2–3 wks): FastAPI/frontend/data compatibility, functional, performance, stress, and regression tests 6. Deployment (1–2 wks): blue-green rollout, monitoring, rollback plan, user docs

Key milestones at weeks 3, 7, 10, 13, and 18.

6. Acceptance Criteria

• Functional parity: agent output quality >= 95% of original; all 7 workflow stages complete; all 6 data sources functional (quality score >= 0.8)
• API compatibility: all existing endpoints (/api/analyze/{symbol}, /api/batch_analyze, etc.) return unchanged formats
• Performance: 100-agent creation >= 100x faster; per-stock analysis <= 30s and >= 50% faster; memory increase <= 500MB and >= 80% lower than original; 10-way concurrency works
• Quality: >= 90% unit, 85% integration, 80% E2E test coverage

7. Risks and Mitigations

High-priority risks: Agno framework immaturity, missing API compatibility, insufficient test coverage. Mitigations include PoC validation, adapter layers, keeping the LangGraph version in parallel, blue-green deployment, and rollback triggers (e.g., abort migration if core features are unachievable or performance is >50% below expectations).

8. Resources and Budget

• Team: 12–14 people (PM, architect, 2 Agno developers, 2 integration engineers, 2 QA, ops, product, QA governance, 1–2 backup)
• Budget: ~$640,000 total ($510k labor, $75k technical, $55k other)
• ROI: estimated $450k/year savings; payback in ~17 months

*Note: figures are the original author's estimates; Agno performance claims are vendor claims, and the Agno code samples are speculative reconstructions.*