Phone-Scale Multi-Agent Debate Loop Deadlocks
This project is a fully local Android-based multi-agent setup that runs four LLM personas in a continuous debate loop using Termux, Ollama, and Llama 3.2 3B Q4_K_M. The personas are deliberately distinct and incompatible: Osmarks is skeptical and analytical, Dominus is dogmatic, Llama is direct and naive, and Satirist keeps tearing down conclusions. The result, according to the post, is not convergence but a stable state of contradiction, all recorded in tamper-evident hash-chained logs and run entirely offline on-device.
Hot take: this is less a demo of “AI debate” and more a proof that role-playing plus incompatible priors can lock a system into productive conflict instead of agreement.
- –The on-device stack is the standout: Termux + Ollama on a Snapdragon 8 Gen 3 makes it genuinely pocket-scale and offline-capable.
- –The persona design appears to be doing most of the work; the system is engineered for epistemic friction, so endless disagreement is a plausible outcome rather than a bug.
- –Hash-chained logs are a smart touch because they make long-running autonomous runs easier to audit and harder to tamper with.
- –This is interesting as an infrastructure experiment, not as a finished product, because the post frames it as a researchy local system rather than something users would directly consume.
DISCOVERED
23d ago
2026-03-20
PUBLISHED
23d ago
2026-03-20
RELEVANCE
AUTHOR
NeoLogic_Dev