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

Description and use

The author built a 17‑agent AI swarm on a Pixel 7, running entirely on-device with TensorFlow Lite, SQLite, and Python threading.

Problem

Running a multi‑agent AI system entirely on a mobile device without cloud or API reliance, while maintaining performance and fault tolerance.

Outcome

The author successfully ran a 17‑agent AI swarm on a Pixel 7, achieving efficient local routing for logistics scenarios without cloud dependency, demonstrating feasibility of on‑device multi‑agent systems.

Workflow

• Train agents on a powerful machine using synthetic map data and reward‑based training.
• Convert Keras models to TensorFlow Lite with post‑training quantization to reduce size and improve speed.
• Deploy quantized models to the Pixel 7 and load them with Python.
• Store map data and agent memory in a local SQLite database.
• Create 17 agents (map sensor, urgency assessment, route evaluation, route coordinator, environmental update) each as a Python thread.
• Implement a central message broker using a Python queue for inter‑agent communication.
• Agents publish outputs as JSON messages to the queue; subscribers consume relevant data.
• Route coordinators aggregate evaluations and select final routes based on urgency, distance, and risk.
• Run the swarm continuously, updating environmental data and re‑evaluating routes as obstacles change.
• Monitor performance and iterate on model size and threading to maintain responsiveness.

Evidence

I’ve managed to get a 17-agent swarm up and running on my Pixel 7, and it’s surprisingly capable.

Source

Training these agents was done on a more powerful machine, naturally.

Source

I used post‑training quantization in TensorFlow Lite:

Source

I used Python's threading module to run each agent in its own thread.

Source