Multi-Agent War Room System

A multi-agent system that simulates a cross-functional "war room" during a product launch. The system analyzes metrics and user feedback to produce structured launch decisions: Proceed, Pause, or Roll Back.

Multi-Agent Orchestration

The system uses LangGraph for workflow orchestration with clear agent separation:

• Data Analyst Agent - Analyzes quantitative metrics, trends, and anomalies
• Product Manager Agent - Defines success criteria and user impact assessment
• Marketing/Comms Agent - Assesses messaging and customer perception
• Risk/Critic Agent - Challenges assumptions and highlights risks
• Coordinator Agent - Makes final decision based on all agent inputs

Agent Flow: Data Analyst → Product Manager → Marketing/Comms → Risk/Critic → Coordinator

Tool Usage

Agents programmatically invoke specialized tools:

• Metric Analysis Tools:
  • analyze_metrics() - Anomaly detection, trend analysis, threshold checking
  • calculate_trends() - Time series analysis and forecasting
• Feedback Analysis Tools:
  • summarize_sentiment() - Sentiment classification and issue categorization
  • extract_common_issues() - Common issue identification and frequency analysis

Setup Instructions

Prerequisites

• Python 3.8+
• AWS Account with Bedrock access (Amazon Nova Pro model)
• AWS SageMaker environment (recommended) or local development
• uv package manager (recommended) or pip

Installation

1. Clone the repository

   git clone https://github.com/adwayapuranik/agent1.git
   cd agent1

2. Install dependencies

   # Using uv (recommended)
   uv sync
   
   # Or using pip
   pip install -r requirements.txt

3. Set up environment variables

   • If using AWS SageMaker, credentials are automatically configured

   export AWS_ACCESS_KEY_ID=your_access_key
   export AWS_SECRET_ACCESS_KEY=your_secret_key
   export AWS_DEFAULT_REGION=us-east-1

4. Configure AWS Bedrock

   • Ensure you have access to Amazon Nova Pro model in AWS Bedrock
   • Model ID: amazon.nova-pro-v1:0

How to Run

End-to-End Execution

Basic run with default dashboard data:

uv run main.py

Run with test scenarios:

# Test proceed scenario (strong positive metrics)
uv run main.py proceed

# Test pause scenario (mixed signals)
uv run main.py pause

# Test rollback scenario (critical issues)
uv run main.py rollback

Example Commands

1. Run with default data:

   uv run main.py

2. Test specific scenarios:

   uv run main.py proceed
   uv run main.py pause
   uv run main.py rollback

3. View logs:

   tail -f data/logs/run_*.log

Input Data Structure (Mock Dashboard)

A) Metrics (data/input/metrics.json)

Time series data for 7-14 days including 6+ key metrics:

• Activation/Signup conversion rate - User onboarding success
• Error rate - API 5xx responses and system failures
• Payment success rate - Transaction completion rates
• Support ticket volume - Customer service load
• DAU/WAU ratio - User stickiness and engagement
• API latency (p95) - Performance metrics in milliseconds

B) User Feedback (data/input/user_feedback.json)

20-50 realistic feedback entries with:

• Date, channel (in-app, support, social), sentiment hints
• User comments covering positive/neutral/negative experiences
• Mix of repeated issues and outlier feedback

C) Release Notes (data/input/release_notes.md)

Feature description for SmartRecommendations v2.1 and known risks

Output Structure (Structured Final Output)

The system produces a structured JSON decision meeting all requirements:

{
  "decision": "Proceed|Pause|Rollback",
  "rationale": {
    "metrics": "Specific metrics analysis driving decision",
    "feedback": "User sentiment impact on decision",
    "technical": "Technical risk assessment",
    "business": "Business impact considerations"
  },
  "risk_register": [
    {
      "risk": "description",
      "severity": "High|Medium|Low", 
      "mitigation": "action plan",
      "owner": "responsible_team"
    }
  ],
  "action_plan": [
    {
      "action": "specific action",
      "timeline": "24h|48h",
      "owner": "responsible_team",
      "success_criteria": "measurable outcome"
    }
  ],
  "communication_plan": {
    "internal": "messaging for teams",
    "external": "customer communication",
    "timeline": "immediate|24h|48h"
  },
  "confidence_score": 0.85,
  "confidence_improvement": "What specific information/actions would increase confidence",
  "next_checkpoint": "when to reassess",
  "escalation_triggers": ["conditions requiring immediate escalation"]
}

Environment Variables Required

Variable	Description	Example
AWS_ACCESS_KEY_ID	AWS access key	AKIA...
AWS_SECRET_ACCESS_KEY	AWS secret key	abc123...
AWS_DEFAULT_REGION	AWS region for Bedrock	us-east-1

Project Structure

agent1/
├── src/agent1/           # Source Layout
│   ├── agents/           # Individual agent implementations
│   │   ├── coordinator_agent.py
│   │   ├── data_agent.py
│   │   ├── marketing_agent.py
│   │   ├── pm_agent.py
│   │   └── risk_agent.py
│   ├── orchestrator/     # LangGraph workflow orchestration
│   │   ├── graph.py
│   │   └── state.py
│   ├── tools/           # Metric and feedback analysis tools
│   │   ├── feedback_tools.py
│   │   └── metrics_tools.py
│   └── utils/           # Logging and utilities
│       └── logger.py
├── data/
│   ├── input/           # Input data files
│   │   ├── metrics.json          # Default dashboard metrics
│   │   ├── user_feedback.json    # Default dashboard feedback
│   │   ├── release_notes.md      # Feature description
│   │   └── examples/             # Test scenario data
│   │       ├── proceed_metrics.json
│   │       ├── proceed_feedback.json
│   │       ├── pause_metrics.json
│   │       ├── pause_feedback.json
│   │       ├── rollback_metrics.json
│   │       └── rollback_feedback.json
│   ├── output/          # Final decisions and analysis
│   │   ├── output.json           # Default scenario output
│   │   ├── proceed_output.json   # Proceed test output
│   │   ├── pause_output.json     # Pause test output
│   │   └── rollback_output.json  # Rollback test output
│   └── logs/           # Execution logs
├── aws/                # AWS Bedrock integration
│   └── bedrock_invoke.py
├── config/             # Configuration files
│   └── config.yaml
├── notebooks/          # Jupyter notebooks
├── main.py            # Main execution script
├── requirements.txt   # Python dependencies
├── pyproject.toml     # uv project configuration
└── README.md          # This file

Traceability & Logging

Complete traceability of agent interactions and tool calls:

• Console output - Real-time agent steps and decisions
• Log files - Detailed execution logs in data/logs/run_YYYYMMDD_HHMMSS.log
• Tool invocations - All tool calls are logged with inputs and outputs
• Agent workflow - Clear progression through Data Analyst → PM → Marketing → Risk → Coordinator

All agent interactions and tool calls are programmatically logged with timestamps.

Example Output

$ uv run main.py

=== RUNNING DEFAULT DASHBOARD ANALYSIS ===
[10:30:15] Data Agent: Analyzing metrics and feedback...
[10:30:18] PM Agent: Assessing user impact and success criteria...
[10:30:21] Marketing Agent: Evaluating customer perception...
[10:30:24] Risk Agent: Identifying critical risks...
[10:30:27] Coordinator: Making final decision...

=== DEFAULT RESULT ===
FINAL DECISION: Pause
Confidence Score: 85%
Key Issues: Mixed signals require investigation before proceeding

Implementation Details

Technology Stack

• Language: Python 3.8+
• Orchestration: LangGraph for multi-agent workflow
• LLM: AWS Bedrock (Amazon Nova Pro model)
• Environment: AWS SageMaker (recommended) or local development
• Package Management: uv (recommended) or pip

Key Features

• Multi-agent coordination with clear role separation
• Programmatic tool usage by agents (not manual analysis)
• Structured JSON output meeting all specification requirements
• Complete traceability of decisions and agent interactions
• Realistic mock data covering multiple decision scenarios

Test Cases

The system includes realistic mock data that demonstrates different decision scenarios:

Default Dashboard Data:

• Run uv run main.py to analyze realistic mixed-signal dashboard data
• Uses data/input/metrics.json and data/input/user_feedback.json
• Demonstrates real-world decision-making with nuanced data

Test Scenarios (data/input/examples/):

• rollback_*: Critical failing metrics → Roll Back decision
• proceed_*: Strong positive metrics → Proceed decision
• pause_*: Mixed signals requiring investigation → Pause decision

Usage:

# Analyze default dashboard data
uv run main.py

# Test specific decision outcomes only
uv run main.py proceed
uv run main.py pause
uv run main.py rollback

Metrics include: 6 key metrics - activation conversion, error rates, payment success, support tickets, DAU/WAU ratio, API latency Feedback patterns: Range from positive to negative sentiment across different channels