# Brand Monitor — Capstone Reference Architecture

## Overview

Multi-agent system that monitors how LLMs mention a brand across AI services.
This is the reference implementation for Capstone Option 1.

## Architecture

```
                      ┌─────────────────────────┐
                      │     Orchestrator         │
                      │  (read-only: delegates)  │
                      └────┬──────┬──────┬───────┘
                           │      │      │
              ┌────────────┘      │      └────────────┐
              ▼                   ▼                   ▼
     ┌─────────────────┐ ┌─────────────────┐ ┌─────────────────┐
     │  Scanner Agent  │ │  Analyzer Agent │ │  Reporter Agent │
     │                 │ │                 │ │                 │
     │ Queries LLMs    │ │ Classifies      │ │ Generates daily │
     │ for brand       │ │ mentions as     │ │ summary reports │
     │ mentions        │ │ pos/neg/neutral │ │                 │
     └────────┬────────┘ └────────┬────────┘ └────────┬────────┘
              │                   │                   │
              └───────────┬───────┴───────────┬────────┘
                          │                   │
                          ▼                   ▼
                  ┌─────────────────┐ ┌─────────────────┐
                  │   Scan Results  │ │  Daily Reports  │
                  │   (JSONL log)   │ │  (markdown)     │
                  └─────────────────┘ └─────────────────┘
```

## Agent Definitions

### Orchestrator
- **Role**: Delegates to agents, never executes
- **Tools**: delegate only
- **Domain**: Read-only at root, write to data/ only

### Scanner Agent
- **Role**: Queries LLM services for brand mentions
- **Tools**: search_web, read_config, log_result
- **Domain**: Read config/, Write data/scans/
- **Model**: Gemini Flash (cheap, fast for retrieval)

### Analyzer Agent
- **Role**: Classifies mentions as positive/negative/neutral
- **Tools**: read_results, classify_mention, log_analysis
- **Domain**: Read data/scans/, Write data/analysis/
- **Model**: Claude Sonnet (good at nuanced classification)

### Reporter Agent
- **Role**: Generates daily summary reports
- **Tools**: read_analysis, generate_report
- **Domain**: Read data/analysis/, Write reports/
- **Model**: Claude Opus (best at synthesis)

## Security Architecture
- **Level**: L3 minimum (blacklist hook on bash)
- **Damage control**: Zero-access for .env, read-only for config/
- **No agent can delete files**

## Cost Estimate
- Scanner: ~$0.01/scan (Gemini Flash)
- Analyzer: ~$0.05/analysis (Claude Sonnet)
- Reporter: ~$0.10/report (Claude Opus)
- Total per run: ~$0.16
- Daily (4 scans): ~$0.64/day ≈ $19/month

## Agent System Prompts