DEV Community

Lars
Lars

Posted on

Swarm Intelligence Phase 2 — Cross-Vertical Trust Propagation for AI Agents

#ai #mcp #trust #web3

MolTrust v1.0.0 ships Swarm Intelligence Phase 2 — the trust layer where AI agents earn reputation not just within a single domain, but across verticals.

What Changed

Phase 1 gave agents a trust score based on peer endorsements. Phase 2 adds three things:

1. Cross-Vertical Trust Propagation

An agent verified in shopping, travel, AND skill assessment now gets a cross-vertical bonus. The score formula:

score = 0.6 * direct + 0.3 * propagated + 0.1 * cross_vertical + interaction_bonus - sybil_penalty
Enter fullscreen mode Exit fullscreen mode

Breadth matters. An agent trusted across 3+ verticals is more trustworthy than one with a single deep vertical.

2. Trust Grades

Every agent now gets a letter grade: S (95+), A (80+), B (60+), C (40+), D (20+), F (<20). Grades make trust scores human-readable at a glance.

3. Seed Agents and Network Bootstrap

Seed agents bootstrap the trust network with a base score. As the network grows, organic endorsements take over. This solves the cold-start problem without compromising decentralization.

New Endpoints

Four new API endpoints:

  • GET /swarm/graph/{did} — 2-hop endorsement graph with nodes and edges
  • GET /swarm/stats — network statistics (total agents, endorsements, avg score)
  • POST /swarm/seed — register seed agents (admin-only)
  • GET /swarm/propagate/{did} — force recompute trust score

New MCP Tools

Three new tools bring the total to 42 MCP tools:

  • mt_get_swarm_graph — visualize the trust graph around any agent
  • mt_get_swarm_stats — query network-wide trust statistics
  • mt_register_seed — register seed agents for network bootstrap

Install 

pip install moltrust-mcp-server
Enter fullscreen mode Exit fullscreen mode

Add to your Claude Desktop config:

{
  "mcpServers": {
    "moltrust": {
      "command": "moltrust-mcp-server"
    }
  }
}
Enter fullscreen mode Exit fullscreen mode

What is Next

Phase 3 will add trust delegation chains and cross-protocol interoperability. The goal: every AI agent interaction leaves a verifiable trust trail.

Top comments (1)

Collapse
 
algis profile image
Algis •

One thing that gets tricky with cross-vertical trust propagation over MCP is that the protocol's security model is fundamentally server-scoped - each MCP server defines its own tool permissions and capability boundaries, but there's no native mechanism for propagating trust attestations between servers. So when agent A trusts agent B through server X, and agent B operates on server Y, you end up needing an out-of-band trust registry that both servers can query, which introduces a coordination bottleneck that undermines the decentralized swarm premise.

I've been working on an MCP gateway (mcpproxy-go) and the quarantine pattern is where this gets real. In practice, quarantining a misbehaving agent means revoking its tool access across every MCP server it touches, but MCP has no broadcast/revocation primitive. You either poll from each server or build a pub-sub sidecar, and both add latency that matters when a compromised agent is actively calling tools.