Peer-to-peer / Swarm

Definition

There is no fixed coordinator. Agents self-organize through direct communication, a shared environment, or dynamic handoff.

Category: Control structure

When to use

Open worlds, decentralized networks, dynamic exploration, research systems, autonomous agent societies.

When not to use

Enterprise production flows, environments with strict permissions, anywhere strong audit and determinism are required.

How to implement

1. Each agent keeps local state and a neighbor list.
2. Use a message protocol; do not depend on a single coordinator.
3. Set TTL, visited set, and budget to prevent flooding.
4. When you need a global result, introduce a temporary aggregator or consensus mechanism.

Minimal pseudocode

async function receive(msg) {
  if (seen(msg.id) || msg.ttl <= 0) return;
  markSeen(msg.id);
  const local = await act(msg);
  for (const peer of pickPeers(local)) {
    send(peer, { ...msg, ttl: msg.ttl - 1, context: local.summary });
  }
}

Recommended trace events

• peer.message.sent
• peer.message.received
• peer.route.selected
• swarm.consensus.reached

Common failure modes

• Hard to converge.
• Duplicated work.
• Weak security boundaries.
• Difficult accountability.

Implementation checklist

• Input/output schemas defined.
• Each agent's permission boundary defined.
• Every agent call carries a run id / trace id.
• Failure, timeout, cancel, and retry strategies defined.
• Context passed is the minimum required, not the full history.
• High-risk actions are gated by approval or a verifier.

References

• Survey of communication