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

Description and use

A 17‑agent swarm on a Pixel 7 uses local LLMs to autonomously gather, analyze, validate, and summarize web3 security trends.

Problem

Run a multi‑agent AI swarm on a mobile device to autonomously gather, analyze, validate, and summarize information about a specific topic.

Outcome

Successfully ran 17 quantized LLM agents on a Pixel 7, achieving decentralized information gathering and summarization with acceptable performance, battery usage, and resilience.

Workflow

• 1. Deploy Redis locally for task queue
• 2. Start coordinator agents to push scrape tasks
• 3. Scraper agents fetch URLs, parse HTML, limit text to 4000 chars
• 4. Analyzer agents perform sentiment, topic extraction, key phrase identification using llama.cpp with Mistral 7B Q4_K_M
• 5. Validator agents cross‑reference data, flag misinformation
• 6. Summarizer agents compile analyzed data into report
• 7. Coordinators collect results, manage retries, handle failures

Evidence

Running 17 LLMs, even quantized ones, on a phone is challenging.

Source

I chose a quantized version of Mistral 7B (specifically Q4_K_M) for its balance of performance and size.

Source

Task Queue (Redis): A Redis server running locally on the phone (using a lightweight Redis implementation) acts as a task queue.

Source

SQLite is used for persistent storage of scraped data, analysis results, and agent logs.

Source