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 locally with TinyLlama 1.1B quantized to 4‑bit, showcasing edge computing and multi‑agent coordination.

Problem

Running a multi‑agent system entirely on a mobile device without cloud dependence, while managing limited CPU, memory, and prompt‑size constraints.

Outcome

The author successfully ran 17 cooperating LLM agents on a Pixel 7, demonstrating that local inference with a small quantized model can support a multi‑agent workflow, albeit with performance and memory trade‑offs.

Workflow

• Select TinyLlama 1.1B and quantize to 4‑bit with llama.cpp
• Implement a lightweight Python agent framework with a shared blackboard
• Define 17 agent roles (Planner, Researcher, Summarizer, Critic, etc.)
• Write generate_prompt to tailor prompts per agent role
• Run agents in a loop: observe → think (LLM generate) → act (update blackboard)
• Optimize prompts for brevity and context window limits
• Manage memory aggressively to avoid OOM errors
• Log and debug interactions via Termux CLI

Evidence

I call it “IronVision Nexus”, and it currently runs 17 distinct agents, each with a specific role, all coordinating (or sometimes arguing) on my Pixel 7.

Source

TinyLlama 1.1B. This is key. While GPT‑3.5/4 would be better quality, they’re simply too large for practical on‑device inference.

Source

I quantized it to 4‑bit using llama.cpp to further reduce memory usage.

Source