Skip to content

agent_framework.md

Latest commit

 

History

History
121 lines (86 loc) · 7.56 KB

agent_framework.md

File metadata and controls

121 lines (86 loc) · 7.56 KB
Raw

Microsoft Agent Framework Investment Workflow

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

Purpose

The Agent Framework implementation is the repository's primary agent-oriented workflow. It uses Microsoft Agent Framework with Azure AI Foundry to coordinate research-only stock-data retrieval, agent-authored Python signal generation, backtesting, plotting, and a final written summary.

Microsoft Agent Framework is the unified successor to AutoGen and Semantic Kernel. It supplies agents for open-ended, tool-using tasks and graph-based workflows for explicit multi-step orchestration. This implementation uses a workflow because the research pipeline has a defined execution order and a conditional backtesting branch.

The workflow produces research artifacts only. It does not provide personalised financial advice, connect to a brokerage, or submit orders.

Architecture

flowchart TD
	classDef orchestrator fill:#f0fff0,stroke:#999,stroke-width:1px
	classDef agent fill:#f0f8ff,stroke:#333,stroke-width:2px
	classDef decision fill:#ffe4e1,stroke:#333,stroke-width:2px

	A[Research request]:::orchestrator --> B[QuantInvestWorkflow]:::orchestrator
	B --> C[WorkflowBuilder]:::orchestrator

	subgraph Pipeline [Type-Safe Workflow Pipeline]
		D[Stock Data Agent]:::agent --> E[Signal Generation Agent<br/>Python REPL]:::agent
		E --> F{signals file?}:::decision
		F -->|exists| G[Backtest Agent]:::agent --> K[Performance Plot Agent]:::agent
		F -->|missing| H[Skip to Summary]:::agent
		K --> I[Summary Report Agent]:::agent
		H --> I
	end
	style Pipeline fill:none,stroke:#333,stroke-width:1px

	C --> D
	I --> J[Final Output]:::orchestrator
Loading

QuantInvestWorkflow constructs the graph with WorkflowBuilder:

Stage Responsibility Implementation
Stock data Retrieves requested OHLCV history. stock_data_fetcher and AgentTools.fetch_stock_data()
Signals Designs a hypothesis, writes Python, executes it, and validates the BuySignal, SellSignal, and Description dataset. signal_generator and AgentTools.run_python_repl()
Conditional route Continues to backtesting only when the validated signal file exists. QuantInvestWorkflow._has_signals()
Backtest Calculates portfolio metrics and writes result artifacts. backtester and AgentTools.backtest_strategy()
Plot Creates the cumulative-return and drawdown chart. performance_plotter and AgentTools.plot_performance()
Summary Writes a bounded research summary with assumptions, limitations, and risks. summary_reporter

The workflow stores checkpoints under checkpoints, emits its Mermaid graph to output/agent_framework/workflow_diagram.mmd after a run, and writes research artifacts below output/agent_framework.

Project modules

Module Role
main.py Loads configuration, creates the workflow, runs a default research request, and optionally launches Dev UI.
workflow.py Defines the Foundry-agent graph and checkpoint integration.
tools.py Implements market-data retrieval, REPL execution, backtesting, plots, and output paths.
models.py Defines Pydantic backtest metrics.

Configuration and run

The root project requires Python 3.13. Install dependencies, copy the template, set the Foundry endpoint and model deployment, then authenticate the Azure CLI:

uv sync
cp .env.example .env
az login
uv run python -m agent_framework.main

On PowerShell, use Copy-Item .env.example .env to create the configuration file.

Variable Default Purpose
AZURE_AI_PROJECT_ENDPOINT — Azure AI Foundry project endpoint. Required.
AZURE_AI_MODEL_DEPLOYMENT_NAME — Azure AI Foundry chat-model deployment. Required.
INVESTMENT_TICKER MSFT Ticker symbol for the research request.
INVESTMENT_START_DATE 2020-01-01 Inclusive market-data start date.
INVESTMENT_END_DATE 2026-07-01 Market-data end date passed to the data provider.
INVESTMENT_INITIAL_CAPITAL 10000 Simulated starting capital.
LAUNCH_DEV_UI false Set to true to serve the workflow in Dev UI on port 8090.

The application calls load_dotenv() itself. Microsoft Agent Framework does not automatically load .env files.

REPL contract and research calculations

The signal agent decides which technical indicators and thresholds to use for each run. It must submit Python to run_python_repl, implemented in research_repl.py, which persists the script as generated_signal_strategy.py and accepts output only when it has one signal row per price row and the required BuySignal, SellSignal, and Description columns. The REPL exposes pandas, NumPy, and ta together with INPUT_PATH and OUTPUT_PATH; it rejects unrelated imports and common dynamic-execution primitives.

The deterministic backtest itself applies the following simplified rules:

  • A position opens on a buy signal only when no position is held, and closes on a sell signal only when a position is held.
  • The return for a held position is the following session's Adjusted Close percentage change, avoiding look-ahead bias.
  • Duplicate buy or sell signals leave the current position unchanged.
  • Metrics include cumulative return, CAGR, maximum drawdown, Sharpe ratio, and final portfolio value.
  • The model does not simulate intraday fills, spread, slippage, trading fees, taxes, or corporate-action validation.

Research artifacts

For the one strategy designed by the agent, the workflow creates:

  • stock_data.csv
  • stock_signals.csv
  • backtest_results.xlsx
  • backtest_metrics.txt
  • stock_plot.png

Human review, patterns, and observability

The primary workflow does not record approvals. For a focused human-review gate, see workflow-checkpointing.py. The standalone pattern library covers agent creation, MCP, RAG, streaming, persistence, retry middleware, structured output, evaluation, observability, and declarative definitions; see Agent Framework patterns.

Console OpenTelemetry exporters are disabled by default in the observability samples. Set ENABLE_CONSOLE_OTEL_EXPORTERS=true only when console traces are needed.

Operational considerations

  • Azure AI Foundry requests use Azure CLI credentials in this implementation; run az login before execution.
  • The workflow executes model-produced code. The REPL validation constrains its intended data contract but is not a security sandbox. Use it only in an isolated development environment with no credentials or production data; replace it with an approved, sandboxed execution service before production use.
  • Use licensed market data and validate data quality, execution assumptions, compliance controls, and human approval requirements before any production deployment.

Official Documentation

Official Microsoft Agent Framework documentation  |  Python samples