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

Description and use

Problem: running a multi‑agent AI system on Android without cloud. Solution: 17 specialized agents running locally via Pyodide, LLaMA.cpp, and 4‑bit Mistral 7B. Result: functional prototype, slow and resource intensive.

Problem

Running a multi‑agent system on a mobile device without cloud dependence, while maintaining privacy and resilience.

Outcome

A proof‑of‑concept 17‑agent AI swarm runs on an Android phone, demonstrating local, distributed intelligence, but is slow, resource intensive, and occasionally crashes the device.

Workflow

• Install Termux and dependencies
• Set up Pyodide in a web view
• Download 4‑bit quantized Mistral 7B model
• Define agent personas in Python
• Implement generate_response and agent_interaction functions
• Create JavaScript sendMessageToAgent to bridge to Pyodide
• Load model and run agents in message‑passing loop
• Handle output flushing with sys.stdout.flush()
• Test agents with sample queries

Evidence

I’ve been obsessively tinkering with running a multi‑agent system – a swarm of 17 individual AI “agents” – entirely on my Android phone.

Source

Pyodide allows you to run Python inside a web browser, using WebAssembly (WASM).

Source

I used a 4-bit quantized version of Mistral 7B.

Source

This project was not easy. Here were the major hurdles: Performance, Memory Management, Pyodide Limitations.

Source