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

Description and use

A 17‑agent AI swarm running entirely on a phone, using a 4‑bit TinyLlama model and lightweight in‑memory queues, demonstrates on‑device AI coordination without cloud reliance.

Problem

Running a swarm of 17 independent AI agents entirely on a mobile device without cloud offloading, while managing thermal, memory, and latency constraints.

Outcome

A 17‑agent swarm that runs entirely on a phone, demonstrating that local, quantized LLMs can coordinate complex tasks without cloud support, albeit with careful resource management.

Workflow

• 1. Select minimal tools: Python, llama-cpp-python, TinyLlama-1.1B-Chat-v1.0 quantized to 4‑bit.
• 2. Define 17 agent roles (Brain, Data Collectors, Analysis, Action Proposal, Evaluation, Execution, Reporting).
• 3. Implement each agent as a Python thread with its own queue interfaces.
• 4. Use in‑memory pub/sub queues for inter‑agent communication.
• 5. Run agents concurrently, monitoring CPU, memory, and temperature.
• 6. Apply throttling to prevent thermal throttling, garbage collection to manage memory, and batching to reduce latency.
• 7. Perform prompt engineering to ensure agents understand roles and cooperate.
• 8. Simulate a market‑prediction scenario to validate swarm behavior.

Evidence

Running 17 concurrent processes, even with a small LLM, is intensive. I had to implement throttling mechanisms to prevent the phone from overheating and shutting down.

Source

I quantized it down to 4‑bit using llama.cpp to further reduce its footprint.

Source

Each agent has its own thread to allow for concurrent operation.

Source