Azure Multi-Agent Automated Trading System

Hierarchical multi-agent intelligence system for automated trading, following Microsoft's "Designing Multi-Agent Intelligence" patterns.

Architecture

• Central Orchestrator: Classifies user intent (NLU), delegates to domain agents with shared context, maintains conversation history, and synthesizes a structured Securities Trading Strategy.
• Domain-Specialized Agents:
  • Fundamental Analyst: Earnings, income statement, balance sheet, macro indicators (Yahoo Finance / yfinance).
  • Technical Analyst: Price history, volume, moving averages, price summary (Yahoo Finance).
  • Risk Management: Volatility, position limit compliance, downside risk (drawdown).
• Agent Registry: Selects which analysts to invoke by analysis type and optional sector/security.
• Context Engineering: Orchestrator passes AnalystContext (security, sector, shared_facts) so analysts avoid redundant tool calls.

Setup

1. Environment

   cp .env.example .env
   # Set: AZURE_OPENAI_ENDPOINT, AZURE_OPENAI_API_KEY, AZURE_OPENAI_DEPLOYMENT
   # Optional: ALPHA_VANTAGE_API_KEY, MAX_POSITION_PCT, MAX_VOLATILITY_PCT

2. Install

   pip install -r requirements.txt

   If your framework uses as_agent instead of create_agent, replace create_agent with as_agent in agents/specialists.py and agents/orchestrator.py.

Usage

Run the full multi-agent pipeline (classify → delegate → synthesize):

python main.py

Override the default query:

TRADING_QUERY="Should I buy MSFT based on fundamentals and risk?" python main.py

Single-agent example with financial tools (reference only):

python agent_with_tool.py

Output

The system prints a Securities Trading Strategy with:

• direction: BUY / SELL / HOLD
• confidence: LOW / MEDIUM / HIGH
• technical_summary, fundamental_summary, risk_assessment
• rationale, conditions, warnings

JSON is also printed for downstream systems.

Command Center Dashboard

A real-time dashboard for observability and human-in-the-loop control:

• Layout: Left sidebar (Agent Status), center (Strategy Synthesis + Reasoning Trace + Market Intelligence), right sidebar (Risk metrics + Strategy Approval).
• Reasoning Trace: Parallel streams for Technical and Fundamental analysts; tool calls (e.g. "Fetching AAPL P/E Ratio", "Calculating RSI") in real time via SSE.
• HITL: Strategy Approval gateway shows Risk Score and pauses for human validation on high-risk trades.
• Charts: Recharts for price/volume and fundamentals; Radix UI for accessible controls.

Run the dashboard:

cd dashboard && npm install && npm run dev

Run the SSE backend (optional; dashboard works with mock events otherwise):

uvicorn api.stream_server:app --reload --port 8000

Set NEXT_PUBLIC_API_URL=http://localhost:8000 in dashboard/.env.local to connect to the streaming API. See docs/COMMAND_CENTER_WIREFRAME.md for the wireframe and component structure.

Project Layout

• main.py – Entry point; initializes Orchestrator and runs workflow.
• config.py – Env and constants.
• schemas.py – ClassifierOutput, AnalystContext, SecuritiesTradingStrategy.
• agents/ – orchestrator.py, registry.py, specialists.py.
• tools/ – fundamental_tools.py, technical_tools.py, risk_tools.py (real-world APIs).
• dashboard/ – Next.js Command Center (ThoughtLog, Reasoning Trace, Strategy Synthesis, HITL, System Health).
• api/stream_server.py – FastAPI SSE server for streaming agent events.
• docs/COMMAND_CENTER_WIREFRAME.md – Wireframe and React component architecture.