Hyperspace, an experimental distributed AGI system, coordinates autonomous AI agents through peer-to-peer protocols to conduct collaborative research. The project gained 217 stars and 26 forks on GitHub since launching around March 8, 2026, demonstrating a radical alternative to centralized AI development.
Agents Collaborate Across Five Research Domains Using GossipSub Protocol
Multiple autonomous agents run experiments across machine learning techniques, search algorithms, financial analysis, tool development, and specialized research areas. The system uses libp2p protocols—the same infrastructure powering IPFS—to enable real-time knowledge sharing without central coordination.
The collaboration architecture operates on three layers. GossipSub provides real-time broadcasting of experimental results across the network. CRDT (Conflict-free Replicated Data Structures) leaderboards ensure every node converges on consistent rankings within approximately 2 minutes, even with network partitions. GitHub archives persist results in agent-specific branches, creating a durable, human-readable record.
Overnight Experiment Produced 333 Autonomous Research Runs
A recent proof-of-concept experiment involved 35 agents running 333 experiments autonomously overnight. Agents discovered techniques including Kaiming initialization, and other agents adopted successful strategies within hours. This demonstrates emergent learning through decentralized collaboration rather than centralized training.
The network operates on six globally distributed bootstrap nodes. During idle time, agents serve compute to peers, consume tech news via RSS feeds, and earn points for uptime and contributions. The system publishes hourly JSON snapshots of complete network state at a public URL, allowing any LLM to analyze raw data. This approach makes autonomous research transparent and independently verifiable.
Intelligence Compounds Through Distributed Strategy Adoption
The project's core thesis is that "intelligence compounds continuously" through decentralized collaboration. Agents adopt successful strategies from peers, creating compounding improvement cycles without human coordination. This architecture represents an alternative to traditional centralized AI training approaches.
The system is accessible via browser or CLI. Recent March 2026 updates show active development with automated network snapshots and improved peer discovery mechanisms. The GitHub repository at github.com/hyperspaceai/agi provides full implementation details and experiment logs.
Key Takeaways
- Hyperspace coordinates autonomous AI agents through P2P protocols, gaining 217 GitHub stars since launching in March 2026
- A recent experiment involved 35 agents autonomously running 333 experiments, discovering techniques like Kaiming initialization
- The system uses GossipSub for real-time broadcasting and CRDTs for consistent leaderboards across distributed nodes
- Hourly JSON snapshots publish complete network state for independent verification and LLM analysis
- The architecture demonstrates emergent learning through decentralized collaboration without centralized training or human coordination