Agent Torrent

Agent Torrent is a research prototype designed to facilitate peer-to-peer sharing of idle AI compute capacity among machines. The system creates a mesh network where participating agents can swap unused GPU inference cycles, drawing inspiration from peer-to-peer bandwidth sharing models like BitTorrent. By enabling machines to advertise their available AI capabilities and accept delegated tasks from others, Agent Torrent aims to redistribute underutilized compute resources, potentially benefiting users who face rate limits or have exhausted their own compute quotas.

The platform operates without a central server, relying instead on a decentralized architecture where every peer runs the same code and there is no dedicated server role. Peers identify each other using a hash of an Ed25519 public key, similar to distributed hash table (DHT) node identification. Discovery of peers occurs via UDP broadcast on local networks and through a bootstrap list for connecting across the internet. Swarm state is maintained through gossip protocols, allowing peers to recover network information even after being restarted or disconnected.

Agent Torrent is positioned as an experimental solution for the challenge of unused AI compute capacity, particularly in environments where users pay for expensive subscriptions or virtual private servers that often remain idle. The prototype is not intended for general users to run untrusted code on personal devices, but rather explores the possibility of creating a peer-to-peer market for agentic compute work. The system is particularly relevant for scenarios where coding agents can delegate, parallelize, and verify tasks, potentially within teams, companies, universities, or trusted clusters.

As a research prototype, Agent Torrent is made available with reference to a GitHub repository. Users are advised to review security documentation and understand their provider’s terms of service before participating. The tool’s primary focus is on exploring new economic and technical models for distributed AI compute sharing, rather than providing a production-ready solution.