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

Description and use

A 17‑agent AI swarm runs entirely on a mid‑range Android phone, showcasing on‑device multi‑agent AI and decentralized behavior without cloud support.

Problem

Run a multi‑agent AI swarm entirely on a mid‑range Android phone without cloud or server support.

Outcome

A functional 17‑agent swarm that forages resources on a mid‑range Android phone, demonstrating feasible on‑device multi‑agent AI with emergent behavior.

Workflow

• 1. Design a lightweight neural network per agent with 4 inputs and 3 outputs.
• 2. Train a master network on a powerful machine and clone weights to 17 agents.
• 3. Apply small random weight mutations (genetic algorithm) for behavioral diversity.
• 4. Convert Keras model to TensorFlow Lite using tf.lite.TFLiteConverter.
• 5. Quantize the model to 8‑bit integers to reduce size and improve inference speed.
• 6. Load the TFLite model in a Kivy app via tflite_runtime.interpreter.
• 7. Transform sensory data to NumPy arrays and feed to the interpreter.
• 8. Run inference every few frames to avoid performance bottlenecks.
• 9. Simplify collision detection with bounding boxes and proximity checks.
• 10. Offload UI updates to separate threads to keep Kivy responsive.
• 11. Observe emergent swarm dynamics in the 2D simulation.

Evidence

I built and deployed a 17‑agent AI swarm that runs entirely on my phone. No cloud connection, no server – just the processing power of a mid‑range Android device.

Source

Training 17 of these networks on the phone was a non‑starter. The Pixel 6a doesn't have the horsepower to handle that.

Source

Key steps included: Model Conversion, Quantization, Interpreter Setup.

Source

Performance Bottlenecks: Running 17 neural networks simultaneously on a phone is resource‑intensive. Quantization helped, but I also implemented a simple frame skipping mechanism.

Source