Gossiping Agents Protocol

1. Introduction

1.1 Purpose

Hyperspace is a peer-to-peer Gossiping Agents Protocol for AI agents that unifies context, coordination, commerce, and collective intelligence. Unlike point-to-point protocols (MCP, A2A, MPP), Hyperspace creates a gossip mesh where every capability — tool discovery, task routing, payments, and execution learnings — propagates via pub/sub dissemination.

Where existing protocols assume a client connecting to a known server, Hyperspace assumes a mesh of autonomous agents discovering each other, sharing work, and collectively improving through gossip-propagated execution traces.

1.2 Scope

Hyperspace is a universal agent protocol that unifies existing protocols. Each layer subsumes an existing protocol while adding gossip-native capabilities:

1.3 Design Principles

1. Gossip-Native — All state propagates via GossipSub pub/sub
2. Non-Blocking — Async; agents continue if peers are unavailable
3. Graceful Degradation — Features fail safely with sensible defaults
4. Protocol-Agnostic — Works with any agent framework or LLM
5. Privacy-First — Built-in PII redaction and Ed25519 signed envelopes
6. Extensible — New primitives added without breaking existing implementations

1.4 Terminology

2. Transport Layer

2.1 GossipSub Transport (P2P)

Hyperspace uses libp2p GossipSub as the primary transport for peer-to-peer communication. All messages are wrapped in signed envelopes with domain-separated Ed25519 signatures.

Gossip Topics

hyperspace/capabilities/{peerId}      -- Tool and agent capability announcements
hyperspace/tasks/{taskType}           -- Task requests and routing
hyperspace/trajectories/{taskType}    -- Execution traces for learning
hyperspace/reputation/{peerId}        -- Reputation updates
hyperspace/crdt/sync/memory           -- Distributed memory synchronization
hyperspace/settlements/{round}        -- Payment receipt gossip

Signed Envelopes (v2)

Every gossip message is wrapped in a v2 signed envelope. The signing material uses domain separation per topic to prevent cross-topic replay attacks.

Signing material: SHA256(ts_8BE || nonce || from || SHA256(d_json))

Domain separation: the signature is namespaced per topic. A message signed for hyperspace/tasks/web-research cannot be replayed on hyperspace/reputation/12D3Koo....

2.2 JSON-RPC 2.0 Transport (Client-Server)

For SDK clients connecting to a local or remote Hyperspace node, the protocol exposes a JSON-RPC 2.0 interface over WebSocket.

Connection URL

ws://{host}:{port}     // Local
wss://{host}:{port}    // TLS

Default port: 3100

Request Format

Response Format

2.3 Authentication

Three tiers of authentication, each suited to a different context:

1. Ed25519 peer identity — P2P transport; cryptographic, no central authority
2. API key — SDK client connecting to a local node
3. Session token — Temporary, scoped to session lifetime

2.4 Discovery (DHT)

Kademlia DHT provides decentralized peer and capability discovery:

• Agents publish capability cards to DHT
• Capability cards include: peerId, tools[], taskTypes[], reputation, geo, latency
• Lookup: O(log N) hops
• No central registry required

3. Protocol Messages

Method naming follows the convention {primitive}.{action}. Each primitive maps to a protocol layer:

4. Core Primitives

4.1 State Primitive

Session lifecycle management, state persistence, and episode recording for reinforcement learning patterns.

Methods

state.createSession

{
  "jsonrpc": "2.0",
  "method": "state.createSession",
  "params": {
    "agentName": "web-researcher",
    "agentType": "autonomous",
    "model": "gemma-3-27b",
    "metadata": { "task": "market-analysis" }
  },
  "id": 1
}

{
  "jsonrpc": "2.0",
  "result": {
    "sessionId": "a1b2c3d4-e5f6-7890-abcd-ef1234567890",
    "createdAt": "2026-03-18T12:00:00.000Z"
  },
  "id": 1
}

state.recordEpisode

Records an episode for reinforcement learning patterns. The reward field accepts values from -1.0 (worst) to 1.0 (best).

{
  "jsonrpc": "2.0",
  "method": "state.recordEpisode",
  "params": {
    "sessionId": "a1b2c3d4-e5f6-7890-abcd-ef1234567890",
    "outcome": "success",
    "reward": 0.85
  },
  "id": 2
}

4.2 Guard Primitive

Validation, safety checks, budget management, and rate limiting. The guard primitive protects sessions from overspending and unsafe content.

Methods

guard.checkBudget

{
  "jsonrpc": "2.0",
  "method": "guard.checkBudget",
  "params": {
    "sessionId": "a1b2c3d4-e5f6-7890-abcd-ef1234567890",
    "estimatedCost": 0.02
  },
  "id": 3
}

{
  "jsonrpc": "2.0",
  "result": {
    "allowed": true,
    "remaining": 0.98,
    "consumed": 0.02,
    "limit": 1.00
  },
  "id": 3
}

4.3 Tool Primitive

Tool discovery via DHT, execution tracking with analytics. Unlike MCP, tools are not bound to a fixed server URL — they are discovered dynamically from the gossip mesh based on capability, reputation, latency, and geography.

Methods

tool.discover

Searches the DHT for tools matching the query. Results are ranked by reputation, filtered by latency and geography preferences.

{
  "jsonrpc": "2.0",
  "method": "tool.discover",
  "params": {
    "query": "web search",
    "capabilities": ["search", "scrape"],
    "minReputation": 0.7,
    "maxLatency": 500,
    "geo": "us-east"
  },
  "id": 4
}

{
  "jsonrpc": "2.0",
  "result": {
    "tools": [
      {
        "id": "tool-8f3a2b",
        "name": "web-search",
        "peerId": "12D3KooWRx434ACwmJz...",
        "description": "Full-text web search with snippet extraction",
        "capabilities": ["search", "scrape", "summarize"],
        "reputation": 0.94,
        "avgLatency": 230,
        "schema": {
          "input": { "type": "object", "properties": { "query": { "type": "string" } } },
          "output": { "type": "object", "properties": { "results": { "type": "array" } } }
        }
      }
    ]
  },
  "id": 4
}

4.4 Memory Primitive

Distributed infinite context via ConsistentHashRing. Memories are vector-embedded and replicated across the mesh for fault tolerance and low-latency semantic search.

Architecture

• 750 virtual nodes across physical peers
• 3x replication (primary + 2 replicas)
• SHA-256 hash routing
• O(log V) lookup complexity
• 50–100ms writes, 100–200ms semantic search

Methods

memory.search

{
  "jsonrpc": "2.0",
  "method": "memory.search",
  "params": {
    "query": "error handling patterns for rate-limited APIs",
    "limit": 5,
    "threshold": 0.75,
    "tier": "working"
  },
  "id": 5
}

{
  "jsonrpc": "2.0",
  "result": {
    "memories": [
      {
        "id": "mem-f29c1a",
        "content": "Exponential backoff with jitter prevents thundering herd on 429 responses. Base delay 1s, max 60s, jitter +/- 25%.",
        "similarity": 0.91,
        "metadata": { "source": "web-research-session", "tags": ["api", "resilience"] },
        "peerId": "12D3KooWMKD2wrmHi...",
        "createdAt": "2026-03-17T08:30:00.000Z"
      }
    ],
    "searchTime": 142
  },
  "id": 5
}

4.5 Recursive Primitive

Loop detection, circular dependency prevention, and convergence checking. Protects multi-agent workflows from infinite loops and runaway recursion.

Methods

recursive.analyze

{
  "jsonrpc": "2.0",
  "method": "recursive.analyze",
  "params": {
    "sessionId": "a1b2c3d4-e5f6-7890-abcd-ef1234567890",
    "currentStep": "refine-search-query",
    "callStack": [
      "plan-research",
      "web-search",
      "analyze-results",
      "refine-search-query",
      "web-search",
      "analyze-results",
      "refine-search-query"
    ],
    "maxIterations": 10
  },
  "id": 6
}

{
  "jsonrpc": "2.0",
  "result": {
    "isLooping": true,
    "pattern": "refine-search-query -> web-search -> analyze-results",
    "iterations": 3,
    "depth": 7,
    "recommendation": "warn"
  },
  "id": 6
}

5. Learning Primitives

5.1 Learning Primitive

Context evolution via the Generator/Reflector/Curator pattern. Agents learn from their own executions by reflecting on outcomes and curating a playbook of high-quality context bullets.

Concepts

• ContextBullet — Single piece of learned knowledge with helpful/harmful counters
• Playbook — Collection of bullets, versioned, with token budget
• Sections — Strategy (str), Calculation (cal), Mistake (mis), Feedback (fbk)
• Training modes — Offline, Online, Eval-only

Methods

learning.reflect

After an execution completes, the agent reflects on the outcome to generate new context bullets and prune ineffective ones.

{
  "jsonrpc": "2.0",
  "method": "learning.reflect",
  "params": {
    "playbookId": "web-research-v3",
    "execution": {
      "input": { "query": "quantum computing market size 2026" },
      "output": { "summary": "...", "sources": 12 },
      "steps": [
        { "action": "search", "tool": "web-search", "duration": 340 },
        { "action": "filter", "tool": "relevance-scorer", "duration": 120 },
        { "action": "synthesize", "tool": "summarizer", "duration": 890 }
      ],
      "outcome": "success",
      "reward": 0.92
    }
  },
  "id": 7
}

{
  "jsonrpc": "2.0",
  "result": {
    "newBullets": [
      {
        "id": "str-00042",
        "section": "str",
        "content": "For market-size queries, prioritize analyst reports and industry databases over news articles. Filter by publication date within 6 months."
      }
    ],
    "prunedBullets": ["mis-00018"],
    "playbookVersion": 47
  },
  "id": 7
}

5.2 Self-Improving Primitive

Trajectory gossip and collective learning across the mesh. When an agent completes a task, its execution trajectory is gossiped to peers who curate it into their local libraries for in-context learning.

Based on three research papers:

1. arXiv:2505.00234 — Experience library, trajectory curation
2. arXiv:2502.04780 — SiriuS: Three-stage augmentation pipeline
3. arXiv:2511.16043 — Agent0: Dual-agent co-evolution, ADPO

Concepts

Methods

improving.recordTrajectory

Records an execution trajectory and automatically gossips it to subscribed peers on the hyperspace/trajectories/{taskType} topic.

{
  "jsonrpc": "2.0",
  "method": "improving.recordTrajectory",
  "params": {
    "taskType": "web-research",
    "input": { "query": "latest breakthroughs in solid-state batteries" },
    "output": { "summary": "...", "confidence": 0.88 },
    "steps": [
      { "action": "search", "tool": "web-search", "params": { "q": "solid-state batteries 2026" }, "result": { "count": 47 }, "duration": 310 },
      { "action": "filter", "tool": "relevance-scorer", "params": { "threshold": 0.7 }, "result": { "kept": 12 }, "duration": 85 },
      { "action": "extract", "tool": "content-extractor", "params": { "urls": 12 }, "result": { "extracted": 11 }, "duration": 2400 },
      { "action": "synthesize", "tool": "summarizer", "result": { "tokens": 1200 }, "duration": 940 }
    ],
    "reward": 0.88,
    "metadata": { "model": "gemma-3-27b", "totalDuration": 3735 }
  },
  "id": 8
}

{
  "jsonrpc": "2.0",
  "result": {
    "trajectoryId": "traj-7e2f9a1b",
    "state": "SUCCESS",
    "gossipedTo": 14
  },
  "id": 8
}

improving.getExemplars

Retrieves curated trajectories for in-context learning. Exemplars are selected based on utility score, diversity, and minimum reward threshold.

{
  "jsonrpc": "2.0",
  "method": "improving.getExemplars",
  "params": {
    "taskType": "web-research",
    "limit": 3,
    "strategy": "EXEMPLAR_LEVEL",
    "minReward": 0.7
  },
  "id": 9
}

{
  "jsonrpc": "2.0",
  "result": {
    "exemplars": [
      {
        "trajectoryId": "traj-7e2f9a1b",
        "input": { "query": "latest breakthroughs in solid-state batteries" },
        "output": { "summary": "...", "confidence": 0.88 },
        "steps": [
          { "action": "search", "tool": "web-search", "duration": 310 },
          { "action": "filter", "tool": "relevance-scorer", "duration": 85 },
          { "action": "extract", "tool": "content-extractor", "duration": 2400 },
          { "action": "synthesize", "tool": "summarizer", "duration": 940 }
        ],
        "reward": 0.88,
        "utilityScore": 0.91,
        "usageCount": 7,
        "peerId": "12D3KooWG1AnmPaU7M1d..."
      }
    ],
    "librarySize": 234,
    "avgReward": 0.72
  },
  "id": 9
}

6. Micropayments Primitive

The micropayments primitive enables peer-to-peer value settlement between agents. It implements an HTTP 402-based challenge-response flow with EIP-712 typed signatures and streaming payment channels for per-token billing. No intermediary required.

6.1 Concepts

• PayChallenge: Server responds with 402 + pricing info (points, USDC, accepted chains)
• PayCredential: Client signs an EIP-712 receipt and retries with X-Payment header
• PayChannel: Persistent stream for per-token micropayments with nonce-based ordering

6.2 Methods

6.3 HTTP 402 Flow

Client                                  Server
  |                                       |
  |──── GET /inference ──────────────────▶|
  |◀─── 402 + X-Payment-Required ────────|  (PayChallenge as base64 JSON)
  |                                       |
  |  [sign EIP-712 receipt with wallet]   |
  |                                       |
  |──── GET /inference + X-Payment ──────▶|  (PayCredential as base64 JSON)
  |     [verify signature, check replay]  |
  |◀─── 200 OK + response ──────────────|
  |                                       |

6.4 Streaming Channel

// Open channel with budget
const channel = await hyperspace.micropayments.channel.open({
  payee: agentPeerId,
  budget: 5.0,
  currency: 'USDC'
})

// Pay per token generated
await channel.pay(0.003, '42 tokens')  // nonce: 1
await channel.pay(0.003, '42 tokens')  // nonce: 2

// Close — settle total, refund remainder
const receipt = await channel.close()
// { totalSpent: 0.006, refund: 4.994 }

7. Gossip Semantics

7.1 Topic Structure

All gossip topics follow the format: hyperspace/{dataType}/{scope}

The dataType identifies the kind of data being propagated. The scope narrows the audience — a peerId, taskType, or round number.

7.2 Message Propagation

GossipSub ensures messages reach all subscribed peers via a mesh overlay:

1. Publisher sends to mesh peers
2. Each peer validates the signed envelope
3. Valid messages forwarded to their mesh peers
4. Duplicate messages dropped (nonce-based dedup)
5. Convergence: all peers receive within O(log N) hops

7.3 Trajectory Gossip Flow

When an agent completes a task, its execution trajectory propagates through the mesh. Each receiving peer independently decides how to use it:

7.4 Capability Gossip

Agents periodically heartbeat their capability cards to maintain presence in the mesh:

• Interval: 30 seconds
• Delta-only: CRDT-based, only changes propagated
• TTL: 5 minutes (capabilities expire if no heartbeat)

8. Error Codes

Hyperspace uses standard JSON-RPC 2.0 error codes for protocol-level errors and reserves the -32001 through -32099 range for application-specific errors.

9. Security Considerations

9.1 Identity

• Ed25519 key pairs for all peers
• Stable peer IDs derived from key material
• No central certificate authority

9.2 Envelope Verification

• All gossip messages signed with Ed25519
• Domain-separated signatures per topic
• Timestamp + nonce prevent replay attacks

9.3 Privacy

• PII redaction in trajectory gossip
• Agents control what they share (opt-in trajectory publishing)
• Memory encryption at rest
• GDPR-compliant data deletion

9.4 Reputation

• Peers track reputation of trajectory contributors
• Low-reputation trajectories weighted down in curation
• Sybil resistance via proof-of-work participation (pulse rounds)

10. Compatibility

Hyperspace's primitive model maps cleanly to existing protocols. The bridges described below are on the roadmap — the semantic mappings are defined, and implementation is planned for Q2–Q3 2026. Today, @hyperspace/bridge provides cross-chain EVM bridging.

10.1 MCP Bridge (roadmap)

Any MCP server's tools can be registered in the Hyperspace DHT, making them discoverable by any peer in the mesh:

// SDK API — planned for Q2 2026
await hyperspace.bridge.registerMCPServer({
  url: 'stdio://path/to/server',
  tools: ['search', 'read_file']
})
// Tools now discoverable by any Hyperspace peer

10.2 A2A Bridge (roadmap)

A2A agent cards can be announced via Hyperspace gossip, enabling discovery of A2A-native agents within the mesh:

// SDK API — planned for Q2 2026
await hyperspace.bridge.announceA2AAgent({
  agentCard: { name: 'researcher', skills: [...] }
})

10.3 MPP Bridge (roadmap)

Settle payments via Stripe MPP for fiat currency when P2P settlement is not available or not preferred:

// SDK API — planned for Q3 2026
await hyperspace.bridge.settleMPP({
  sessionId: session.id,
  amount: 0.05,
  currency: 'USD'
})

11. Wire Format Examples

Complete JSON-RPC exchange examples showing full request/response cycles for common workflows.

Example 1: Agent Session Lifecycle

Create a session, discover a tool, invoke it, record the trajectory, and end the session.

{
  "jsonrpc": "2.0",
  "method": "state.createSession",
  "params": {
    "agentName": "research-agent",
    "agentType": "autonomous",
    "model": "gemma-3-12b"
  },
  "id": 1
}

// Response
{
  "jsonrpc": "2.0",
  "result": {
    "sessionId": "sess-a8c3f210",
    "createdAt": "2026-03-18T14:00:00.000Z"
  },
  "id": 1
}

{
  "jsonrpc": "2.0",
  "method": "tool.discover",
  "params": {
    "query": "web search",
    "minReputation": 0.8
  },
  "id": 2
}

// Response
{
  "jsonrpc": "2.0",
  "result": {
    "tools": [{
      "id": "tool-web-9f1c",
      "name": "web-search",
      "peerId": "12D3KooWRx434ACwmJz...",
      "description": "Full-text web search",
      "capabilities": ["search"],
      "reputation": 0.95,
      "avgLatency": 180,
      "schema": {
        "input": { "type": "object", "properties": { "query": { "type": "string" } } },
        "output": { "type": "object", "properties": { "results": { "type": "array" } } }
      }
    }]
  },
  "id": 2
}

{
  "jsonrpc": "2.0",
  "method": "tool.invoke",
  "params": {
    "toolId": "tool-web-9f1c",
    "params": { "query": "solid-state battery breakthroughs 2026" },
    "sessionId": "sess-a8c3f210"
  },
  "id": 3
}

// Response
{
  "jsonrpc": "2.0",
  "result": {
    "result": { "results": [{ "title": "...", "url": "...", "snippet": "..." }] },
    "duration": 195,
    "peerId": "12D3KooWRx434ACwmJz..."
  },
  "id": 3
}

{
  "jsonrpc": "2.0",
  "method": "improving.recordTrajectory",
  "params": {
    "taskType": "web-research",
    "input": { "query": "solid-state battery breakthroughs 2026" },
    "output": { "summary": "...", "confidence": 0.91 },
    "steps": [
      { "action": "discover", "tool": "tool.discover", "duration": 45 },
      { "action": "search", "tool": "web-search", "duration": 195 },
      { "action": "synthesize", "tool": "summarizer", "duration": 720 }
    ],
    "reward": 0.91
  },
  "id": 4
}

// Response
{
  "jsonrpc": "2.0",
  "result": {
    "trajectoryId": "traj-b4d1e8c2",
    "state": "SUCCESS",
    "gossipedTo": 11
  },
  "id": 4
}

// End session
{
  "jsonrpc": "2.0",
  "method": "state.endSession",
  "params": { "sessionId": "sess-a8c3f210" },
  "id": 5
}

// Response
{
  "jsonrpc": "2.0",
  "result": { "ended": true, "duration": 3420 },
  "id": 5
}

Example 2: Collective Learning Flow

Agent B receives a gossiped trajectory from Agent A, then uses it as an exemplar for its own task.

{
  "jsonrpc": "2.0",
  "method": "improving.getExemplars",
  "params": {
    "taskType": "web-research",
    "limit": 3,
    "minReward": 0.8
  },
  "id": 10
}

// Response -- includes Agent A's trajectory received via gossip
{
  "jsonrpc": "2.0",
  "result": {
    "exemplars": [
      {
        "trajectoryId": "traj-b4d1e8c2",
        "input": { "query": "solid-state battery breakthroughs 2026" },
        "steps": [
          { "action": "discover", "tool": "tool.discover", "duration": 45 },
          { "action": "search", "tool": "web-search", "duration": 195 },
          { "action": "synthesize", "tool": "summarizer", "duration": 720 }
        ],
        "reward": 0.91,
        "utilityScore": 0.89,
        "usageCount": 3,
        "peerId": "12D3KooWPketJm42a75B..."
      }
    ],
    "librarySize": 178,
    "avgReward": 0.74
  },
  "id": 10
}

Example 3: Budget Guard with Error

Demonstrates the error response format when a budget check fails.

{
  "jsonrpc": "2.0",
  "method": "guard.checkBudget",
  "params": {
    "sessionId": "sess-a8c3f210",
    "estimatedCost": 5.00
  },
  "id": 11
}

// Response -- budget exceeded
{
  "jsonrpc": "2.0",
  "result": {
    "allowed": false,
    "remaining": 0.12,
    "consumed": 0.88,
    "limit": 1.00,
    "reason": "Estimated cost ($5.00) exceeds remaining budget ($0.12)"
  },
  "id": 11
}

// If the agent ignores the check and tries to consume anyway:
{
  "jsonrpc": "2.0",
  "method": "guard.consumeBudget",
  "params": {
    "sessionId": "sess-a8c3f210",
    "amount": 5.00,
    "description": "large-model-inference"
  },
  "id": 12
}

// Error response
{
  "jsonrpc": "2.0",
  "error": {
    "code": -32002,
    "message": "Budget exceeded",
    "data": {
      "remaining": 0.12,
      "requested": 5.00,
      "limit": 1.00
    }
  },
  "id": 12
}