# Demo Video Scenario — Saiten Scoring Agent
> Recording guide for the Agents League @ TechConnect submission demo.
> Target length: **3–4 minutes**.
---
## Pre-Recording Checklist
- [ ] VS Code open with this workspace
- [ ] GitHub Copilot Chat panel visible (right sidebar)
- [ ] Terminal panel closed or minimized
- [ ] `data/scores.json` reset or empty (for clean demo)
- [ ] MCP server running (auto-starts via `.vscode/mcp.json`)
---
## Scene 1: Introduction (0:00 – 0:30)
**Visual**: VS Code with README.md open, Copilot Chat panel visible.
**Narration / Caption**:
> "Saiten is a multi-agent scoring system for the Agents League hackathon.
> It uses 6 Copilot custom agents and an MCP server to automatically
> collect, score, review, and report on 43+ submissions — all from
> VS Code's chat panel."
**Action**: Briefly scroll through the README showing the workflow diagram.
---
## Scene 2: Scoring a Single Submission (0:30 – 1:30)
**Prompt to type in Chat**:
```
@saiten-orchestrator score #49
```
**Expected Flow** (show each step appearing in chat):
1. **Orchestrator routes** → delegates to `@saiten-collector`
2. **Collector** fetches Issue #49 (EasyExpenseAI)
- Shows: Track, repo URL, README length, demo status
- Gate: ✅ Data complete
3. **Scorer** evaluates with evidence
- Shows per-criterion scores with evidence citations
- Example: "Accuracy: 8/10 — 5-agent Semantic Kernel pipeline"
4. **Reviewer** validates
- Shows: PASS (within 2σ, no generic phrases)
5. **Reporter** generates report
- Shows: `reports/ranking.md` saved
**Narration / Caption**:
> "Each agent has a single responsibility. The scorer cites specific
> evidence for every criterion — no generic 'good README' phrases allowed."
---
## Scene 3: Scoring All Submissions (1:30 – 2:30)
**Prompt to type in Chat**:
```
@saiten-orchestrator score all
```
**Expected Flow**:
1. **Orchestrator** starts processing 43 submissions
2. **Todo list** appears in VS Code showing progress
- "Collecting submissions..." → ✅
- "Scoring Creative Apps (30)..." → ✅
- "Scoring Reasoning Agents (10)..." → ✅
- "Scoring Enterprise Agents (3)..." → ✅
- "Reviewing scores..." → ✅
- "Generating report..." → ✅
3. **Final output**: Top 10 table with scores
4. **Handoff button** appears: "💬 Post feedback comments to Top 10"
**Narration / Caption**:
> "The Evaluator-Optimizer pattern ensures scoring consistency:
> the reviewer checks for bias and statistical outliers,
> and flags submissions for re-scoring if needed."
---
## Scene 4: Review & Report (2:30 – 3:15)
**Action**: Open `reports/ranking.md` to show the generated report.
**Show**:
- Top 10 overall ranking table
- Per-track champion sections (Creative, Reasoning, Enterprise)
- An individual evaluation summary showing evidence-anchored scores
**Prompt (optional)**:
```
@saiten-orchestrator review scores
```
**Expected output**:
- Bias detection results (5 types checked)
- Score distribution statistics
- Any flagged outliers
**Narration / Caption**:
> "The reviewer agent detects 5 types of bias: central tendency,
> halo effect, leniency, range restriction, and anchoring.
> Flagged scores get sent back for re-evaluation automatically."
---
## Scene 5: Human-in-the-Loop & MCP (3:15 – 3:45)
**Action**: Click the "💬 Post feedback comments to Top 10" Handoff button.
**Show**:
- `@saiten-commenter` generates comment previews
- User confirms before any GitHub Issue comment is posted
- Comment includes score breakdown and improvement suggestions
**Narration / Caption**:
> "Comments are only posted after explicit user confirmation —
> the Human-in-the-Loop pattern ensures no accidental feedback."
**Action**: Show `.vscode/mcp.json` briefly to demonstrate zero-config setup.
---
## Scene 6: Testing & Closing (3:45 – 4:00)
**Action**: Run in terminal:
```bash
python -m pytest tests/ --cov=saiten_mcp -q
```
**Show**: `100 passed, 93% coverage` output.
**Narration / Caption**:
> "100 automated tests with 93% code coverage verify
> every parser, model, tool, and report generator."
---
## Key Points to Highlight
| Feature | What to Emphasize |
| ----------------------------- | --------------------------------------------- |
| **6 Custom Agents** | Orchestrator-Workers pattern with SRP |
| **MCP Server** | 5 tools, FastMCP, stdio transport |
| **Evidence-Anchored Scoring** | Every criterion requires cited evidence |
| **Evaluator-Optimizer** | Reviewer → re-score feedback loop |
| **Human-in-the-Loop** | Handoff for comment posting |
| **100 Tests / 93% Coverage** | pytest + Pydantic validation |
| **Zero Config** | `.vscode/mcp.json` auto-starts MCP |
| **Copilot Usage** | Entire agent system built with GitHub Copilot |
---
## Fallback: If Demo Breaks
If the MCP server or GitHub API is slow, you can:
1. Show `data/scores.json` with pre-scored results
2. Show `reports/ranking.md` with the generated report
3. Run `python -m pytest tests/ -q` to demonstrate testing
4. Walk through the agent `.md` files to explain the architecture
