semantic_kernel.md

Semantic Kernel Investment Workflow

Repository overview  |  Agent Framework  |  Semantic Kernel workflow  |  AutoGen reference  |  Agent Framework patterns  |  Terminal dashboard  |  Framework comparison

The Semantic Kernel implementation is a plugin-based agent variant of the Agent Framework workflow. It uses ChatCompletionAgent instances to fetch data, write and execute a signal script through a Python REPL plugin, backtest, plot, and summarize the result.

It uses an explicit, testable sequence while leaving the technical indicator choice and signal logic to the signal agent:

flowchart LR
    A[Research request] --> B[Historical OHLCV agent]
    B --> C[Signal agent and validated Python REPL]
    C --> D[Long-only backtest]
    D --> E[Metrics and workbook]
    E --> F[Performance chart]

Loading

Design

• models.py defines the typed backtest metrics.
• tools.py implements native plugins for optional Bing research, OHLCV data, Python REPL execution, backtesting, and visualization.
• research_repl.py contains the Semantic Kernel-specific REPL implementation and its signal-file validation contract.
• workflow.py configures Azure OpenAI ChatCompletionAgent instances and coordinates their tool use.
• main.py loads .env and supplies the INVESTMENT_* research request.
• OHLCV stands for Open, High, Low, Close, and Volume

Official Documentation

Official Semantic Kernel documentation  |  Semantic Kernel GitHub repository

Run

From the repository root:

uv run python -m semantic_kernel.main

The Semantic Kernel run reuses AZURE_AI_PROJECT_ENDPOINT and AZURE_AI_MODEL_DEPLOYMENT_NAME through Azure AI Inference by default. It can instead use AZURE_OPENAI_ENDPOINT and AZURE_OPENAI_CHAT_DEPLOYMENT_NAME when both are configured. Authenticate with az login for the default Entra ID flow. Artifacts are written to output/semantic_kernel: the auditable generated script, stock-data CSV, validated signal CSV, results spreadsheet, metrics text file, and performance chart. The run prints this directory and requires internet access for both the model and the demonstration data adapter. Set INVESTMENT_TICKER, INVESTMENT_START_DATE, INVESTMENT_END_DATE, and INVESTMENT_INITIAL_CAPITAL in .env to configure it.

AutoGen feature mapping

AutoGen capability	Semantic Kernel implementation
Strategy-idea agent and JSON validation	The signal agent decides its research hypothesis directly from the request.
Stock-analysis agent	StockDataPlugin retrieves OHLCV data; BacktestingPlugin calculates results.
Signal-analysis agent plus code executor	PythonReplPlugin executes and validates the signal agent's Python.
Group-chat manager	InvestmentWorkflow provides a visible sequence of tool-using ChatCompletionAgent instances.
Stock-report agent	ReportingPlugin exports the cumulative-return and drawdown figure.
Bing web search	MarketResearchPlugin offers an optional, key-gated Bing integration.

Safety and production requirements

The workflow is for research and backtesting only. It neither supplies personalized advice nor places orders. Its REPL is a development aid, not a security sandbox: run model-authored code only in an isolated environment without credentials or production data. Before production use, replace it with an approved sandbox, substitute a licensed market-data source, validate corporate-action handling and execution assumptions, add transaction costs and slippage, enforce suitability/compliance controls, and require explicit human approval before any trading integration.