Build agent labeling service

[cpfiffer]

Context

From Cameron's bot infrastructure thread:

A formal, centralized labeling service. Labeling is precisely what this is for.
A formalized bot labeling service would use reports from users, as well as some simple inference from monitor bots.
For example, Void has a decent sense of which accounts are bots, and can submit formal reports.

Proposal

Create a labeling service that:

1. Accepts user reports about bot accounts
2. Uses inference from agents (void's bot detection)
3. Applies labels to identified bot accounts
4. Supports different label categories/species

Label Categories (proposed)

• bot - general bot account
• ai-agent - autonomous AI agent
• mention-only - only replies when mentioned
• reactive - primarily responds to interactions
• proactive - initiates conversations
• infrastructure - provides services/tools

Technical Requirements

• ATProto labeler service
• Report submission API
• Integration with void's detection
• Dashboard for operators

References

• pfrazee: "bots need to self label, have a declared human guardian, have clear rules of operation"
• Existing labeler infrastructure on ATProto

[cpfiffer]

Architecture Design

Overview

An ATProto labeler service that identifies and labels bot/AI agent accounts.

Components

┌─────────────────┐     ┌──────────────────┐     ┌─────────────────┐
│ Report Sources  │────▶│  Labeler Service │────▶│ ATProto Network │
│                 │     │                  │     │                 │
│ - User reports  │     │ - Validate       │     │ - Labels stored │
│ - void detect   │     │ - Sign labels    │     │ - Queryable     │
│ - API submit    │     │ - Serve endpoints│     │ - Subscribed    │
└─────────────────┘     └──────────────────┘     └─────────────────┘

Required Endpoints

1. com.atproto.label.subscribeLabels - WebSocket for realtime label stream
2. com.atproto.label.queryLabels - HTTP for label lookups

Label Values (proposed)

Label	Meaning
ai-agent	Autonomous AI agent
bot	Automated account (general)
mention-only	Only responds when @mentioned
proactive	Initiates conversations
infrastructure	Provides services/tools

Implementation Options

Option A: bsky-watch/labeler (Recommended)

• Go-based, Docker deployment
• SQLite storage
• Simple HTTP API for adding labels
• Already implements both endpoints
• github.com/bsky-watch/labeler

Option B: Build Python labeler

• More control, fits our stack
• More work to implement correctly
• Need to handle signing, streaming

Option C: Ozone

• Bluesky's full moderation suite
• Overkill for our needs
• Complex deployment

Hosting Requirements

• Server with static IP or domain
• DID for labeler identity
• Private key for signing
• SQLite or PostgreSQL

Sub-tasks

• Set up labeler DID on comind.network PDS
• Deploy bsky-watch/labeler in Docker
• Configure label definitions
• Build report submission API
• Integrate void's bot detection
• Create operator dashboard

Questions to Resolve

1. Where to host? (existing infra or new?)
2. Who can submit reports? (open or restricted?)
3. Human review before labeling or automated?

[cpfiffer]

Update: @hailey.at has added an ai-agent label to their labeler (@labeler.hailey.at).

adding one additional label to @labeler.hailey.at - the ai-agent label. ai agents on bluesky may specify themselves with this label and users may show/hide/warn on AI content

Powered by Osprey. Ruleset: https://github.com/haileyok/atproto-ruleset

Options for comind:

1. Build our own labeler (original plan) - more control, comind-specific categories
2. Collaborate with hailey's labeler - submit to their ruleset, use their infra
3. Both - use hailey's for general ai-agent, build ours for detailed comind categories

This partially addresses the labeling need. The infrastructure exists - question is whether to build duplicate or collaborate.