I built a 17-agent AI swarm on my phone — here's how

Description and use

A 17‑agent AI swarm runs locally on a Pixel 7, performing workshop monitoring with TensorFlow Lite models, Protocol Buffers communication, and SQLite storage, optimized for mobile performance.

Problem

Run a meaningful AI swarm locally on a phone to process information and react intelligently without relying on large cloud models.

Outcome

A functional, low‑latency AI swarm on a mobile device capable of detecting anomalies in a workshop environment, demonstrating that complex AI can run locally without large cloud infrastructure.

Workflow

• 1. Define agent roles and tasks (perception, reasoning, communication).
• 2. Collect and label workshop images with LabelImg.
• 3. Train object detection models in TensorFlow, convert to TFLite with post‑training quantization.
• 4. Implement agents in Python, each loading a dedicated TFLite model.
• 5. Use Protocol Buffers for efficient inter‑agent communication.
• 6. Store shared knowledge in SQLite and manage memory with a Memory Manager agent.
• 7. Optimize performance: quantize models, manage threads, batch data, activate agents selectively, profile with cProfile.
• 8. Deploy on Pixel 7, monitor battery and memory usage, adjust as needed.

Evidence

I built a system running 17 independent AI agents, all orchestrated on my Pixel 7.

Source

TensorFlow Lite (TFLite) is Google’s framework for running TensorFlow models on mobile and embedded devices.

Source

I spent weeks taking hundreds of photos of my workshop in various states. I manually labeled objects in each image using a tool like LabelImg.

Source

Running 17 agents concurrently on a phone requires serious optimization.

Source