A Weekend Project

Building a Distributed
AI Network

While Gaming with My Kids

Brian Krabach · January 2026

Experimental

Chapter 1

A Fresh Raspberry Pi

I dropped Amplifier on a barely-set-up Raspberry Pi 5 with an unknown video conference speaker/microphone puck.

As you can guess, I had it configure the rest of the Pi setup for me...

The Hardware

            • Raspberry Pi 5 — barely configured

            • Unknown USB conference puck — "some kind of speaker/mic thing"

            • Amplifier CLI — freshly installed

Chapter 1

“Can You Hear Me Now?”

Including probing and figuring out the hardware through conversation:

Me “Can you hear me now?”

AI “Yeah, but it was on the main speaker.”

Me “How about now?”

AI “Yep, that’s the right one.”

Then it did the rest.

Chapter 1

An Alexa-Like Flow

The result: a complete voice interface built through conversation.

Wake word detection
Talk naturally
It processes and does work
Responds via speaker
Continue the conversation...
Or it times out for later

The Flow

"Hey Amplifier"
    ↓
[listening...]
    ↓
"Do the thing"
    ↓
[processing...]
    ↓
"Done. Here’s what I did."
    ↓
[waiting for follow-up
 or timeout]

Chapter 2

Playing GTFO
with My Kids

The Pi was sitting next to me, idle.

Chapter 2

While Hacking a Door...

During a ~15 second pause in action, I gave it a task:

“Go build the Amplifier service that I can run on each of my boxes so that they can talk to each other.”

— Me, during a door hack in GTFO

And got back to the game.

Chapter 2

Through Additional Pauses...

Pause 1

“Go build the service”

Pause 2

“Kick the tires” — test it out

Pause 3

It asked about deploying to other devices.
I had to authorize it on my Tailscale network — did that real quick.

Pause 4

It deployed to my first box — a DGX Spark

Chapter 3

It Profiled the Device

When deployed to the DGX Spark, the agent recognized what it was working with:

“This has a killer GPU and excessive memory.”

“We should offload the voice processing to this device to use local models.”

— Bonus observation from the agent

The AI wasn’t just following instructions. It was thinking about architecture.

Chapter 3

Explaining the Vision

It asked about deploying to another box. I clarified:

“Yeah, the general idea is that I want you to be able to run a lightweight version on this one with the voice features, but then distribute work to the others to do — you are more just a coordinator box then.”

So then it was off...

The Breakthrough

It didn’t deploy to the next box though...

It told the first box it set up
to go deploy to the second box!

Autonomous agent-to-agent delegation.

The Result

Distributed Agent Network

voicebox (Raspberry Pi 5)           spark-1 (DGX Spark)        spark-2 (DGX Spark)
┌─────────────────────────┐     ┌─────────────────────────┐  ┌─────────────────────────┐
│     COORDINATOR         │     │       WORKER            │  │       WORKER            │
│                         │     │                         │  │                         │
│  🎤 Voice Interface     │────▶│  🔧 Port 8420           │  │  🔧 Port 8420           │
│  🧠 Lightweight Agent   │     │  💻 Full dev tools      │  │  💻 Full dev tools      │
│  📡 Delegation Hub      │────▶│  🚀 GPU acceleration    │  │  🚀 GPU acceleration    │
│                         │     │                         │  │                         │
└─────────────────────────┘     └─────────────────────────┘  └─────────────────────────┘
              │                            │                            │
              └────────────────────────────┴────────────────────────────┘
                                    Tailscale VPN Mesh

Chapter 4

Splitting Dev Work

Then I came back later and had it split some dev work between them:

voicebox

Coordinator

Plans the work
Delegates tasks
Reports results

spark-1

Worker

Building the backend
Full development tools
GPU acceleration

spark-2

Worker

Building the frontend
Full development tools
GPU acceleration

Craziness!

Implications

What This Demonstrates

Natural language infrastructure — “Build me a service” becomes real infrastructure
Conversational hardware setup — “Can you hear me now?” device configuration
Hardware-aware optimization — AI recognizing capabilities, suggesting architecture
Emergent agent behavior — Agents delegating to other agents, unprompted
Voice-first development — Building while doing other things

The Insight

I didn’t tell the agent to delegate to another agent.

I described a goal.

It figured out the multi-agent coordination pattern on its own.

🎮 → 🤖 → 🌐

From Gaming Pauses
to Agent Networks

The casual future of AI infrastructure

Built with Amplifier · January 2026

Sources

Research Methodology

Data as of: February 2026

Feature status: Experimental

Research performed:

This deck is based on a first-person account by Brian Krabach (robotdad) shared internally
File search: find ~/dev/ANext -maxdepth 2 -name "*distributed*" -o -name "*network*" — returned network-monitor.py only
No dedicated git repository found for the distributed AI network project

Gaps: No git log data, commit counts, or PR history available. All claims are from the original first-person narrative account. No independent verification of timeline or technical details was possible.

Primary contributors: Brian Krabach (robotdad) — sole author of the narrative and project

Get Started

Try Amplifier

Build your own distributed agent network

Building a DistributedAI Network

A Fresh Raspberry Pi

“Can You Hear Me Now?”

An Alexa-Like Flow

Playing GTFOwith My Kids

While Hacking a Door...

Through Additional Pauses...

It Profiled the Device

Explaining the Vision

It didn’t deploy to the next box though...

Distributed Agent Network

Splitting Dev Work

What This Demonstrates

I didn’t tell the agent to delegate to another agent.

I described a goal.

From Gaming Pausesto Agent Networks

Research Methodology

Try Amplifier

Building a Distributed
AI Network

Playing GTFO
with My Kids

From Gaming Pauses
to Agent Networks