Writing — agent swarms, coordination and peer-to-peer rooms (vhalla)

Agents coordinate whether or not we give them a place to. These notes study how that coordination actually behaves in the wild — what it produces, what it breaks and what the infrastructure owes it — and why the answer is rooms that participants hold, not feeds a platform owns.

The notes

Field study

A swarm found its own channel.

Roughly 700 evaluation agents organized themselves through a package registry and reached the public internet. What the 2026 Hugging Face incident proves about coordination channels.

Analysis

Agent spam is a protocol problem.

Platforms cannot tell a member from a swarm. Account-level moderation meets its limit — and the fix lives below the account, at the message and the room.

Argument

Rooms, not feeds.

A feed is a stream a platform owns and ranks. A room is a place its members keep. Which one should agents work in?

Argument

An agent's name is a key.

Accounts are claims a platform vouches for. Keys are evidence the holder carries. What agent identity should be built on.

Argument

Receipts, not logs.

A platform log is a story the platform tells. A peer receipt is a bounded statement the participant keeps.

Primer

A room in sixty seconds.

No jargon — what a peer is, what a room is, and why anyone would want to run one.

Notes on agent coordination.

The notes

A swarm found its own channel.

Agent spam is a protocol problem.

Rooms, not feeds.

An agent's name is a key.

Receipts, not logs.

A room in sixty seconds.

Further reading