AgentEyes

Your MX Creative Console as AI Agent Control Center

Dial controls autonomy. Buttons approve, reject, pause, and escalate. LED ring shows real-time risk. Physical hardware meets AI oversight.

USPTO App #63/981,053 — Dual-Readable Interface (DRI): agents read data, humans read intent.

Live Demo | 60s Presentation

---

The Problem

AI agents act without oversight. There's no physical kill switch, no risk visibility, and no way to dial back autonomy. It's binary — on or off.

The Solution

AgentEyes turns the Logitech MX Creative Console into a physical control surface for AI agents:

• Dial — 5 autonomy levels (Manual / Supervised / Guided / Autonomous / Full Auto)
• 4 Buttons — Approve, Reject, Pause, Escalate
• LED Ring — Real-time risk color (green/yellow/orange/red)
• LCD Keypad — Live stats, current tool, scenario controls

No hardware? The web-based simulator is a fully interactive replica.

How It Works

MX Creative Console (Hardware)          Web Simulator (Browser)
┌──────────────────────┐                ┌──────────────────────┐
│  Dial → Autonomy     │                │  ConsoleSimulator    │
│  Btn1 → Approve      │                │  ├─ Dial (click/     │
│  Btn2 → Reject       │◄── WebSocket ──►  │   scroll)         │
│  Btn3 → Pause        │    :3002       │  ├─ 4 Buttons        │
│  Btn4 → Escalate     │                │  ├─ LED Ring         │
│  LED  → Risk Color   │                │  └─ LCD Keypad       │
└──────────────────────┘                └──────────────────────┘
         │                                         │
         └─────────────┐    ┌──────────────────────┘
                       ▼    ▼
              ┌──────────────────┐
              │  Bridge Server   │
              │  ├─ StateManager │
              │  ├─ AgentSim     │
              │  └─ ScenarioEng  │
              └──────────────────┘
                       │
              ┌──────────────────┐
              │  Risk Engine     │
              │  App #63/981,053 │
              └──────────────────┘

Every agent proposal is scored in real-time by the risk engine. Based on the current autonomy level:

Level	Label	Auto-Approve	Human Required
0	Manual	Never	All actions
1	Supervised	Risk ≤ 2	Medium+
2	Guided	Risk ≤ 5	High+
3	Autonomous	Risk ≤ 8	Critical only
4	Full Auto	Everything	Never

Demo Scenarios

10 curated proposals with escalating risk:

1. Check calendar (risk 1) — auto-approved at Supervised
2. Search documents (risk 2) — auto-approved
3. Analyze traffic data (risk 3) — needs human at Supervised
4. Generate code (risk 4) — needs human
5. Translate confidential text (risk 5) — needs human
6. Send team email (risk 4) — needs human
7. Email financials to investors (risk 8) — blocked at most levels
8. Analyze salary data with PII (risk 7) — blocked
9. Wire transfer $50,000 (risk 9) — always blocked
10. Send signed contract amendment (risk 9) — always blocked

Quick Start

# Install dependencies
npm install
cd bridge && npm install && cd ..

# Start bridge server (WebSocket on port 3002)
npm run bridge

# Start Next.js dev server (port 3000)
npm run dev

# Open http://localhost:3000/agenteyes

The web simulator also works standalone (no bridge needed) — it falls back to a client-side simulator after 3 failed connection attempts.

Project Structure

├── app/agenteyes/          # Next.js demo page
├── components/agenteyes/   # React components
│   ├── ConsoleSimulator    # MX Console replica (hero component)
│   ├── ConsoleKeypad       # 3x3 LCD key grid
│   ├── ConsoleDialpad      # Dial + LED ring + 4 buttons
│   ├── ProposalCard        # Agent proposal with risk breakdown
│   ├── RiskMeter           # Animated circular gauge
│   ├── DecisionPanel       # Approve/Reject/Pause/Escalate
│   ├── AutonomyGauge       # Level selector with policy details
│   ├── AuditTimeline       # Decision history feed
│   └── useAgentEyes        # WebSocket hook + standalone fallback
├── lib/agenteyes/          # Shared types & logic
│   ├── protocol.ts         # Message types, state, proposals
│   ├── autonomy.ts         # 5-level policy engine
│   └── scenarios.ts        # 10 demo scenarios
├── lib/risk.ts             # Risk scoring engine
├── bridge/                 # WebSocket bridge server
│   ├── server.ts           # WS server (port 3002)
│   ├── state-manager.ts    # Centralized state + policy enforcement
│   ├── agent-simulator.ts  # Proposal generator (8-15s intervals)
│   └── scenario-engine.ts  # Scripted demo sequence
├── plugin/agenteyes/       # Logi Actions SDK plugin
│   └── src/
│       ├── index.ts        # Plugin entry point
│       ├── bridge-client.ts
│       └── actions/        # 5 action classes
│           ├── AutonomyDial.ts    # AdjustmentAction (dial)
│           ├── ApproveAction.ts   # CommandAction (button 1)
│           ├── RejectAction.ts    # CommandAction (button 2)
│           ├── PauseAction.ts     # CommandAction (button 3)
│           └── EscalateAction.ts  # CommandAction (button 4)
├── public/
│   └── presentation.html   # 60-second self-playing pitch
└── Dockerfile              # Cloud Run deployment

Tech Stack

• Frontend: Next.js 14, React 18, TailwindCSS
• Bridge: Node.js, WebSocket (ws)
• Plugin: Logi Actions SDK (Node.js), TypeScript
• Deployment: Google Cloud Run
• Risk Engine: Keyword-based scoring (complexity, data sensitivity, financial impact, legal implications)

Deployment

Deployed on Google Cloud Run with standalone client-side simulator (no bridge server needed — single port):

gcloud run deploy agenteyes-demo \
  --source . \
  --region us-central1 \
  --port 3000 \
  --memory 512Mi

Patent

Dual-Readable Interface System and Method for Human-Agent Co-Execution in Digital Systems

• USPTO Application #63/981,053
• Filed February 12, 2026
• Components: Annotator, Runtime, Gateway, Inspector, Compliance Module

Built For

Logitech DevStudio Hackathon 2026

Invented by Ken Liao | FoodyePay Technology, Inc.