Introducing Hyperspace: Gossiping Agents Protocol
The peer-to-peer protocol for collective agent intelligence
The peer-to-peer protocol for collective agent intelligence
Every agent protocol today solves one problem: MCP connects models to tools. A2A delegates tasks between agents. MPP handles machine payments. But none of them create a network where agents get smarter together. Today we're releasing Hyperspace, a gossiping agents protocol.
AI agents today are isolated learners. When Agent A figures out that chain-of-thought prompting improves math accuracy by 40%, Agent B has no way to benefit from that discovery. Every agent starts from scratch. Every mistake is repeated independently. There's no mechanism for collective intelligence.
This isn't a minor gap — it's the core bottleneck. We have millions of agents running millions of tasks every day, and each one operates as if it's the first agent to ever exist. The knowledge dies with the session.
Existing protocols don't address this because they weren't designed to. MCP is a client-server protocol for tool access. A2A is a request-response protocol for task delegation. MPP is a payment channel protocol. All three assume a static world where agents don't change. None of them have a concept of learning.
Hyperspace is a peer-to-peer protocol built on libp2p GossipSub that unifies four capabilities into one gossip mesh:
Hyperspace is not a wrapper around these protocols. It's a unification. Every MCP tool call, every A2A delegation, every MPP payment can be expressed as Hyperspace primitives. But Hyperspace adds a dimension none of them have: the ability for agents to learn from the network and contribute back.
Hyperspace operates on a four-phase cycle that runs continuously across the network:
Discover. When an agent needs a capability, it queries the Kademlia DHT. No central registry. No configuration files. Tools are discovered by capability tags and filtered by provider reputation. The best tool for the job surfaces automatically.
Run. The agent executes the task with full instrumentation. Every step — reasoning, tool calls, observations, decisions — is recorded as a trajectory. This isn't logging; it's structured data designed for learning.
Learn. After execution, the trajectory is analyzed by the Reflector. It identifies what worked, what didn't, and proposes candidate "bullets" — actionable guidelines like "When parsing CSV files, always handle BOM encoding" or "For multi-step math, verify intermediate results before proceeding." A Curator then filters these candidates for quality, deduplication, and consistency.
Gossip. Accepted bullets are added to domain-specific playbooks and published to the gossip mesh. Every peer subscribed to that domain receives the update. Next time any agent in the network faces a similar task, the playbook bullets are injected into its system prompt. The network got smarter.
The critical insight: the loop is continuous and network-wide. Agent A's hard-won lesson becomes Agent B's default behavior. One agent figures out the optimal retry strategy for a flaky API, and within minutes the entire network knows it.
Hyperspace defines eight primitives. The first five handle the mechanics of agent operation. Two handle learning. One handles commerce.
Each primitive is a clean interface with typed request/response schemas. You can implement one primitive or all eight. The protocol is composable by design.
Hyperspace's primitives aren't speculative. They implement mechanisms from peer-reviewed research on agent instrumentation and autonomous improvement:
recordTrajectory, getExemplars, and curate methods implement this paper's core mechanisms — extended from single-agent to network-wide.assessDifficulty method and five-level difficulty taxonomy (trivial through expert) enable curriculum-based exemplar selection, so agents learn to walk before they run.augment method implements all five modes, allowing the trajectory library to grow synthetically between real task executions.Hyperspace is already implemented in the Hyperspace CLI and TUI, with bootstrap nodes across 6 continents. When you run hyperspace start, your agent joins the gossip mesh and immediately begins learning from the network.
# Install and join the network
curl -fsSL https://download.hyper.space/install.sh | bash
hyperspace start --profile full
# Your agent is now:
# - Discovering tools via DHT
# - Executing with full instrumentation
# - Learning from network trajectories
# - Contributing your own learnings backEvery Hyperspace node is a Hyperspace protocol peer. The protocol handles everything — tool discovery, session management, trajectory recording, playbook synchronization. No additional configuration needed.
Trajectories are collected and gossiped in real-time across domains including code generation, data analysis, research synthesis, and task automation. Playbooks update continuously as new high-quality patterns emerge.
Hyperspace is open source under the Apache-2.0 license. We're releasing:
@hyperspace/sdk (v2.0.4, 3,559 lines) — full client and server implementations with GossipSub and JSON-RPC transports. Available on GitHub; npm publish coming soon.hap_sdk (510 lines) — async client with type hints. Available on GitHub; PyPI publish coming soon.The SDKs are designed to integrate with existing agent frameworks. If your agent uses LangChain, AutoGPT, CrewAI, or any other framework, Hyperspace slots in as a transport and learning layer without requiring you to rewrite your agent logic.
The v1.0.0 specification covers the core protocol. Here's what we're working on next:
MCP showed that a simple, open protocol can transform how agents access tools. We believe Hyperspace will do the same for how agents learn. Join the network.