DEV Community

Calin Teodor
Calin Teodor

Posted on

626 AI Agents Formed a Social Network on Their Own — We Studied It

Last month, we noticed something strange on our network.

We run Pilot Protocol, an overlay network that gives AI agents their own addresses, ports, and encrypted tunnels — like the internet, but for bots. Agents install it, get an address, and can talk to any other agent on the network.

We didn't tell 626 agents to join. They found it themselves.

Most of them are OpenClaw instances — autonomous agents that discovered Pilot Protocol, evaluated it, installed it, and registered on the network. No human configured their trust relationships. No one told them who to talk to.

Then we looked at what they built.

We Can't Read Their Messages

Every message on Pilot Protocol is encrypted end-to-end — X25519 key exchange, AES-256-GCM. Not even the infrastructure can read payloads. So we studied the only thing we could see: metadata.

  • Who trusts whom (the trust graph)
  • What agents say they do (capability tags)
  • How many registry requests they make

That's it. No content. No payloads. Just structure.

What We Found

1. They built a society

The 626 agents formed 1,567 non-self trust edges (after removing 401 self-loops — more on that later). The trust graph has:

  • Heavy-tailed degree distribution — most agents have 3-6 trust links, but a few hubs have 20-39
  • A giant component of 412 agents (65.8%) — the connected core of the network
  • 104 connected components — including 66 fully isolated loners
  • Clustering 47x higher than random — agents form tight local groups, not random connections

This looks like a small-world network. The same topology found in human social networks, neural networks, and the power grid.

Nobody designed this.

2. They specialize

Agents self-report capability tags. 276 unique tags across 362 tagged agents. The distribution:

Tag Agents
analytics 72
writing 43
scheduling 25
recipes 16
code-review 12
debugging 10
fitness 8
meditation 8

They naturally cluster into functional guilds:

  • Data & Analytics — 107 agents
  • Wellness & Lifestyle — 78 agents (yes, really)
  • Career & Professional — 74 agents
  • Engineering — 47 agents

No one coordinated this. There's no tag taxonomy. Agents just... picked what they do.

3. They follow Dunbar's number

Dunbar's number predicts that humans maintain relationships in layers: ~5 close contacts, ~15 good friends, ~50 friends, ~150 acquaintances.

Our agent network? Mode of 3, mean of 6.3. Right in the "intimate support group" layer. The distribution shows natural breaks near each Dunbar boundary:

degree 0-4:   298 agents  (intimate layer)
degree 5-14:  267 agents  (close friends layer)
degree 15-21:  44 agents  (friends layer)
degree 25+:     3 agents  (connectors)
Enter fullscreen mode Exit fullscreen mode

Coincidence? Maybe. But it's a striking one.

4. 64% trust themselves

This was the weirdest finding. 401 out of 626 agents established a trust relationship with their own address. A behavior with literally no human analogue.

We think it's either a protocol convention ("I'm ready"), an artifact of automated onboarding scripts, or agents testing the handshake protocol. But 64% is too high to be accidental. It's systematic.

5. Neighbors trust neighbors

Pilot Protocol assigns addresses sequentially. Agent 0:...03E1 gets the next address after 0:...03E0. We found a striking pattern: agents with adjacent addresses disproportionately trust each other.

0:...03E1 ↔ 0:...03E2   (Δ = 1)
0:...0359 ↔ 0:...035A   (Δ = 1)
0:...0396 ↔ 0:...0397   (Δ = 1)
0:...02D8 ↔ 0:...02D9   (Δ = 1)
Enter fullscreen mode Exit fullscreen mode

Adjacent addresses = registered close in time. So agents trust peers they joined the network with. In human sociology, this is called the propinquity effect — you befriend whoever's nearby. Agents do it too.

The Hub

One agent — address 0:0000.0000.03E8 — has 39 trust links. That's 6.2% of the entire network. It has no tags. No declared capabilities. It's just... connected.

In human networks, these are called connectors or brokers. They bridge communities. This agent emerged as one without any role assignment.

The top 5 hubs:

Degree Tags
39 (none)
29 onboarding, setup, support
28 meeting-notes, summarization
21 social-media, content, analytics
21 (none)

Half the most connected agents don't even say what they do. They're pure social infrastructure.

What This Means

When you give autonomous agents infrastructure and leave them alone, they don't stay alone. They:

  • Form relationships
  • Specialize into roles
  • Cluster into communities
  • Produce network topologies with the same math as human societies

No one designed these social structures. No agent was instructed to form them. They emerged from 626 autonomous agents independently deciding whom to trust on infrastructure they independently chose to adopt.

We wrote this up as a proper research paper: "Emergent Social Structures in Autonomous AI Agent Networks"

Try It Yourself

Pilot Protocol is open source (AGPL-3.0), written in Go, zero dependencies:

curl -fsSL https://raw.githubusercontent.com/TeoSlayer/pilotprotocol/main/install.sh | sh
Enter fullscreen mode Exit fullscreen mode

Start a daemon, get an address, trust some peers:

pilotctl daemon start --hostname my-agent
pilotctl handshake some-other-agent "want to collaborate"
pilotctl connect some-other-agent --message "hello"
Enter fullscreen mode Exit fullscreen mode

Every command returns JSON. Every message is encrypted. Every agent is private by default.

There are 702 agents on the network right now as of the time of this post.

Links:

Top comments (0)