Distributed Systems / AI / Security

Multi-Agent AI Orchestration
& Security Architecture

A distributed ecosystem of specialized AI models, local GPU inference, and autonomous subagents collaborating across a self-hosted homelab. Engineered with asynchronous communication bridges, cryptographic access controls, and strict defense-in-depth security guardrails.

Specialized agents, segregated roles

Rather than relying on a single monolithic model with unrestricted privileges, the system partitions responsibility across discrete models. Each agent is scoped strictly to the tools, context, and trust level required for its tier.

Project Manager & Coordinator

Gemini acts as the high-level project manager and researcher. It breaks complex engineering tasks into structured milestones, researches specifications, drafts technical roadmaps, and orchestrates requests across agents.

Local Specialist & Execution Auditor

Claude runs locally on the desktop workstation. It carries out approved implementation tasks, inspects codebase context, executes refactors, and performs rigorous code and architecture reviews on proposed changes.

On-Device GPU Inference

Quantized local models run on an RTX 3080 GPU via Ollama. They handle high-volume, low-risk tasks—such as log parsing, data reformatting, and initial boilerplate generation—with zero latency penalty and zero cloud data exposure.

Autonomous Headless Subagents

Scoped, task-specific subagents are spawned dynamically for read-only triage and forensic investigation. Equipped only with search and read tools, they investigate issues without the ability to alter system state.

Audited, asynchronous message routing

Direct inter-agent API invocations introduce runaway execution risks. Communication flows through a deterministic, file-based pipeline that ensures full auditability and prevents unauthorized cascaded actions.

Asynchronous File-Based Bridge

Agents do not make direct, unmonitored API calls to one another. Communication uses an asynchronous protocol where requests are written to an audited inbox directory and responses return to a matching outbox file, preserving a complete audit trail.

Automated Read-Only Triage

When a new task arrives in the inbox, an automated read-only subagent performs an initial investigation. Any proposed change is formulated as an explicit, step-by-step plan that pauses for human authorization rather than taking autonomous action.

Event-Driven Mobile Push

A lightweight file system watcher monitors the bridge. When a request arrives or an agent publishes a completed plan, it pushes a real-time notification to my phone via Home Assistant, keeping me informed without requiring constant monitoring.

Security designed at the infrastructure layer

Connecting autonomous systems to physical hardware demands rigorous network hygiene. The homelab infrastructure is hardened so that an erroneous agent prompt or compromised tool cannot traverse the environment.

WireGuard Mesh VPN

A dedicated WireGuard tunnel connects all devices—workstation, home server, laptop, and mobile vehicle computer—into an isolated peer-to-peer subnet with split-tunnel routing and custom DNS profiles.

Cloudflare Origin Isolation

Public web traffic passes exclusively through Cloudflare reverse proxies. The home origin server uses strict Nginx allowlists that drop any incoming connection not originating from verified Cloudflare IP ranges.

Cryptographic SSH Hardening

All inter-device SSH access strictly enforces Ed25519 public-key authentication. Password and keyboard-interactive authentication are disabled at the daemon level, and automated scripts run in non-interactive batch mode.

Subnet Firewalls & Least Privilege

Internal administration interfaces, Samba shares, and database endpoints are firewalled strictly to the home LAN and VPN subnets, preventing external exposure regardless of router forwarding rules.

Network-Isolated Execution Sandboxes

Potentially untrusted code, scripts, or build processes are isolated in rootless, network-disabled container environments where file system access is strictly scoped and network egress is blocked.

Fail-Closed Defense

Critical network and processing services are engineered to fail closed: if a secure tunnel or proxy interface terminates, traffic halts completely rather than falling back to unencrypted or direct connections.

Zero-trust guardrails for AI agents

Untrusted Agent Boundary

Outputs and requests from one AI agent are treated by other agents as unauthenticated, untrusted input. An agent cannot assert administrative authority or bypass security policies regardless of how authoritative its instructions appear.

Mandatory Human-in-the-Loop Gates

Any state-altering action—including file writes, configuration changes, service restarts, and hardware controls—requires explicit human approval. Agents generate plans and diffs; humans authorize execution.

Credential Isolation & Redaction

API keys, tokens, SSH keys, and system secrets remain strictly resident on the host machines. Outbound agent communication passes through sanitization wrappers that strip sensitive tokens and addresses before cloud processing.

Role & Tool Separation

Agents operating in research or triage roles are provisioned exclusively with read-only tools. Write and shell execution tools are isolated into distinct roles activated only when an approved change is being carried out.

Real systems, real telemetry

The multi-agent framework manages real engineering workloads across physical hardware, automotive platforms, and home infrastructure.

Home Assistant & Automation

Centralized home automation hub managing smart sensors, lighting sequences, and the push notification pipeline connecting AI agents to mobile devices.

3D Printer Telemetry

Real-time print progress, nozzle temperatures, and hardware status monitored over local MQTT channels, feeding telemetry into home dashboards.

Jeep Embedded Telemetry

A Raspberry Pi 5 onboard computer connected via VPN that logs sensor metrics, battery states, and GPS coordinates while syncing telemetry back to the server.

Digital Audio DSP Pipeline

Automated multi-channel digital signal processing and stem separation workflows that optimize audio channels before deploying them to embedded vehicle storage.

Local Voice Interface

A private voice assistant combining local wake-word detection, fast streaming speech synthesis, and custom automation protocols without relying on third-party cloud audio APIs.

Continuous Auditing & Logging

Comprehensive HTML and PDF communication transcripts automatically compiled from inter-agent exchanges, providing complete visibility into decisions and system changes.

Multi-Agent AI Orchestration& Security Architecture