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

---

💸 Investment Agent Patterns

This project shows how to use Microsoft Agent Framework for stock-market research. Agents download past prices, write and execute a technical-analysis signal script, test the resulting strategy, and save the results.

Real-time terminal dashboard and backtesting interface

chart-cli renders a live watchlist, price chart, and Agent Framework workflow output in the terminal. Press / for the command prompt: /ask answers questions about the symbols on screen, /backtest has an agent write and test a strategy from a plain-language request, /period sets the window, /model switches the chat backend, and /help lists everything. In the screenshot below, AGENT:RULES means the workflow is using its transparent rule-based path because no chat provider was connected.

Run this from PowerShell:

cd chart-cli
pnpm install
.\scripts\test-cli.ps1

The script verifies the Node UI and Python Agent Framework protocol, then opens the interactive dashboard in the same terminal. It starts the workflow process used by the dashboard; do not start a separate Python backend. Use .\scripts\test-cli.ps1 -CheckOnly for verification without opening the UI.

See the terminal dashboard guide for the panels, keys, commands, periods, and provider setup.

What's included

Area	Purpose	Use
Agent Framework workflow	Main workflow.	Start here.
Semantic Kernel workflow	Same steps, plugin-based.	Compare with the main workflow.
AutoGen reference	Earlier group-chat version.	Compare implementation styles.
Agent Framework patterns	Examples for the investment domain.	Explore features one at a time.
Real-time terminal dashboard	Live watchlist and charts driven by an Agent Framework workflow, with /ask, /backtest, /period, and /model commands.	Watch signals update in a terminal.
Framework comparison	Short comparison of the three frameworks.	-

The Agent Framework patterns are the only pattern showcase in this repository. The 30 examples show Microsoft Agent Framework features for investment research. Semantic Kernel and AutoGen do not have separate pattern libraries here.

Quick start: Microsoft Agent Framework

Python 3.13 and uv. To run the main Agent Framework example, you also need an Azure AI Foundry project with a chat model and access through the Azure CLI.

uv sync
cp .env.example .env   # set Azure AI Foundry endpoint and model deployment
az login                # authenticate the Azure CLI credential used by the workflow
uv run python -m agent_framework.main

On PowerShell, use Copy-Item .env.example .env to copy the settings file. Then fill in these values in .env:

Variable	Purpose
AZURE_AI_PROJECT_ENDPOINT	Address of the Azure AI Foundry project used by the main workflow.
AZURE_AI_MODEL_DEPLOYMENT_NAME	Name of the chat model deployed in that project.
AZURE_OPENAI_ENDPOINT	Optional direct Azure OpenAI endpoint for the Semantic Kernel workflow; otherwise it uses the Foundry project endpoint.
AZURE_OPENAI_CHAT_DEPLOYMENT_NAME	Optional direct Azure OpenAI deployment for the Semantic Kernel workflow.
INVESTMENT_TICKER	Stock symbol to study. The default is MSFT.
INVESTMENT_START_DATE, INVESTMENT_END_DATE	First and last dates for the past-price data.
INVESTMENT_INITIAL_CAPITAL	Pretend starting amount for the backtest.

Optional: Semantic Kernel variant

Use this command only when comparing the plugin-based agent variant. It reuses the configured Foundry project and model by default, then creates charts, metrics, spreadsheets, CSVs, and the generated signal script in its own output folder:

uv run python -m semantic_kernel.main

Sample research output

The following Agent Framework sample uses MSFT from 2020-01-01 through 2026-07-01 with a simulated $10,000 starting balance. It is historical research only; a later run can generate a different strategy and result.

Input

Analyze MSFT from 2020-01-01 to 2026-07-01. Develop one transparent technical-analysis signal strategy as Python code, execute it, backtest $10000, and report CAGR, total return, final value, drawdown, and Sharpe ratio.

Generated signal

The sample agent generated a long-only trend-and-momentum signal using MSFT closing prices:

• Buy: the 50-day simple moving average is above the 200-day simple moving average and RSI(14) is above 50, after that combined condition was previously false.
• Sell: the combined condition becomes false: the 50-day average is at or below the 200-day average, or RSI(14) is at or below 50.
• Hold: no change in the combined condition.
• Execution convention: a close-based condition affects the next session's return, reducing same-close look-ahead bias.

Performance plot

The top panel shows cumulative strategy returns. The lower panel shows portfolio drawdown from its previous peak.

Backtest metrics

Metric	Sample value
Cumulative return	27.96%
CAGR	3.88%
Maximum drawdown	-21.03%
Sharpe ratio	0.36
Final value	$12,796.05

See the full backtest workbook, generated signal script, and validated signals.

Repository layout

Path	Contents
agent_framework	The main Microsoft Agent Framework application.
semantic_kernel	The Semantic Kernel version of the research workflow.
autogen	The AutoGen reference version, with its own Poetry environment.
agent_framework_patterns	Thirty small Agent Framework examples for investment research.
chart-cli	Real-time terminal dashboard with a companion Agent Framework workflow.
tests	Offline tests for Agent Framework patterns and the REPL contracts.
output	Saved charts, metrics, and example pattern responses.
docs	Guides for each implementation and framework comparison.

Validation

The Agent Framework pattern tests run without Azure credentials or live market-data services:

uv run ruff check agent_framework semantic_kernel agent_framework_patterns tests
uv run pytest tests -q

Safety and limitations

• These results are for learning and research, not financial advice or a real trading system.
• Both workflows execute model-authored Python to create signals. Their validation is not a security sandbox; run them only in an isolated development environment without credentials or production data.
• Good results from the past do not mean the same strategy will work in the future.
• The examples use public market data and simple rules. They leave out trading fees, price changes that happen while a trade is being made, taxes, careful handling of stock splits and dividends, and checks for an individual investor's needs.
• Review AI responses, connected tools, and data licences before using this project outside a learning or research setting.

📝 License

MIT