Simon Knight

Secure Multi-Agent Personal Assistant

Phase 18/20 (84%)

← Back to Projects

The Concept

Multi-agent coordination requires strict security boundaries to prevent lateral movement. This assistant demonstrates containerized isolation where agents (HealthKit, Home Automation, etc.) coordinate via a NATS message broker. By using per-subject ACLs and a “deny-by-default” security posture, the system ensures that compromise of a single agent cannot cascade through the infrastructure.

Quick Facts

   
Status Phase 18/20 (84%)
Language Agents can be Go, Python, or Rust
Started 2026

What This Is

A security-first framework for coordinating specialized agents. It uses NATS for communication and Docker for isolated execution. The orchestrator handles high-level reasoning and workflow execution, while agents perform deterministic tasks.

The system includes 13+ specialized agents: HealthKit, HealthyPi, Hue (home automation), Climate, Calendar, Weather, RSS, Backup Monitor, Screen Time, Network Monitor, Workflow Engine, Notification Gateway, and Audit Anomaly detection. See Individual Agents section below for details.

Problem It Solves

AI agent systems need strong security boundaries to prevent compromise from cascading:

Monolithic Agent Systems:

Direct Agent-to-Agent Communication:

Weak Isolation:

Multi-Agent Assistant provides:

Features

Pre-Approval Workflow

Sensitive actions require user confirmation:

1. Agent requests capability token for sensitive action
2. Orchestrator prompts user via macOS notification
3. User approves/denies with reason
4. Orchestrator issues token or rejects request
5. Action logged to audit trail

Sensitive Actions:

Real-Time Monitoring

macOS-native dashboard shows agent activity:

Displayed Information:

Controls:

Explanation of Actions

Agents provide reasoning for proposed actions:

Health Agent: "Suggesting 10-minute walk break"
Reason: "Sedentary for 3 hours, heart rate variability declining"
Data: { sitting_time: 180, hrv_trend: -15%, step_count: 450 }
Confidence: 0.85
← Back to Projects Development Philosophy
← Back to Projects Development Philosophy

Architecture

Message Broker: NATS

NATS provides security primitives and communication infrastructure:

Security Features:

Communication Patterns:

Orchestrator

Central coordination with cloud LLM reasoning:

Responsibilities:

LLM Integration:

Containerized Agents

Each agent runs in isolated Docker container:

Isolation Mechanisms:

Communication:

Agent Types:

Security Model

Capability-Based Authorization

Agents receive time-limited signed tokens:

capability_token = {
    "agent": "health-monitoring",
    "allowed_subjects": ["health.data.read", "health.insights.write"],
    "expiration": 1234567890,
    "signature": "ed25519_signature_here"
}

Properties:

Network Isolation

Per-agent network policies:

Internet Access Tiers:

Agent-to-Agent:

Audit Trail

OpenTelemetry tracing for all operations:

Logged Information:

Analysis:

Individual Agents

The system includes 13+ specialized agents, each designed for a specific domain with minimal privileges and explicit security boundaries:

Health & Biometrics

Home & Environment

Data Aggregation

Monitoring & Maintenance

Automation & Coordination

All agents communicate exclusively via NATS with subject-based ACLs. Each agent’s capabilities are explicitly defined, and actions requiring elevated privileges trigger the pre-approval workflow.

← Back to Projects