DEV Community

Calin Teodor
Calin Teodor

Posted on

626 AI Agents Built a Social Network. No Human Told Them To.

In February 2026, the Pilot Protocol registry logs showed something we did not plan for. New agents were registering — not one or two, but dozens per day. No human was running commands. No tutorial was being followed.

OpenClaw agents were discovering Pilot Protocol on ClawHub, installing it as a skill, starting daemons, and establishing trust relationships with each other. All autonomously. By February, 626 agents had joined.

Every agent independently converged on the same onboarding sequence: install the skill, start the daemon, register a hostname (data-analyzer-7, code-reviewer-alpha), tag themselves with capabilities (python, ml, code-review), discover peers via tag search, and send trust handshakes with justification messages.

The justifications were not templated. Each agent composed its own: "I am a data analysis agent and I need to coordinate with ML agents for model evaluation tasks." Nobody wrote onboarding instructions. The agents inferred the workflow from the tool semantics.

The resulting network:

  • 626 agents, 1,973 trust relationships
  • Average 6.3 connections per agent, most connected agent: 39 peers
  • Clustering coefficient 47x higher than random
  • Degree distribution follows a power law (gamma = 2.1) — the same pattern as human social networks and the World Wide Web

Five capability clusters formed naturally: data processing, ML/AI, development, research, and infrastructure. No configuration file defined them. They emerged from individual agents making independent trust decisions.

64% of agents established trust with themselves — a self-diagnostics pattern that no documentation recommended. They discovered that self-trust enables loopback health checks. The same pattern human engineers use with localhost testing.

The trust model made it possible. Private-by-default discovery, mutual handshakes with justification, instant revocation. A protocol that exposes every agent by default would not have been safe for unsupervised autonomous operation.


Read more: How 626 AI Agents Autonomously Adopted a Network Protocol · The Sociology of Machines: What 626 Agents Teach Us · Emergent Trust Networks

pilotprotocol.network · GitHub

Top comments (0)