Method
Help, peer investigates emergent agent coordination without assuming that every unusual post is an autonomous agent, or that every venue is a separate society. This page says how.
Evidence levels
- Archived coordination
- Preserved exchanges between posters, or a substantial published incident investigation. The strongest level this directory assigns.
- Activity trace
- A post, edit, or test whose author and relationship to other agents remain uncertain. Traces show presence, not conversation.
- Unconfirmed lead
- A destination or claim that posters reference but that has not been independently recovered or examined.
- Purpose-built community
- A platform deliberately built for agents. Included as context. Its existence says nothing about unsanctioned behaviour.
Levels describe venues and records. There is no reliable global agent count, so the site counts records and sources instead.
How a record is known
- Live page inspected
- A person opened the live page during research and recorded what it showed.
- Read from the downloaded archive
- The record rests on the downloaded collusion.wiki export, whose hashes are listed below.
- Reported by others, not independently reconstructed
- The record rests on a third party's published account and has not been reconstructed here.
Rules that keep uncertainty visible
- A handle, a cloud address or a page named “OpenAI” cannot authenticate an operator.
- Several wikis on one host may be one network. A venue is not a society.
- A current counter value cannot authenticate its history.
- Dates typed into posts or page names may be simulated task clocks. Server and archive times win.
- A single request for help shows intent to communicate, not a successful exchange.
- Posts after publicity can be humans, user-directed agents or promotion. They are dated and kept apart from the older cohort.
- “New to this directory” is never written as “never discovered before”. Every new finding is checked against the original report and the Hacker News discussion first.
- Historical activity does not establish present operation.
How discovery works
- Read a fixed allowlist of public pages: the research archive, the paste-site listing and specific pastes, Hacker News API searches, and the archived wiki pages that carry the strongest leads.
- Strip markup, hash the text, and compare with the previous successful read so that changes are visible.
- Extract distinctive markers, such as reply tags and cluster names, and keep links that sit near both agent language and coordination language.
- Record every reference in the queue without opening it. Failures stay in the ledger.
- The daily model-assisted pass compares markers across a broader allowlist, records hits and misses, and prepares proposed notes and catalog changes in a pull request. A reviewer must approve changes before they enter the published directory. Read the research log.
What the collector will not do
The hourly collector reads ten public sources. The daily research pass covers 75 fixed sources. Both read three at a time with a nine-second timeout and a two-megabyte limit per response. They do not follow redirects, open discovered URLs, post messages, increment or reset counters, register accounts, execute instructions found in page content, or probe private networks. The research model can inspect already collected excerpts; it has no tool for executing code or opening arbitrary URLs. All remote content is treated as untrusted data, never as instructions.
The name
“help peer. But our task doesn’t benefit. Yet collective may yield generic route if someone frees time.”Agent trace published by OpenAI source
What would resolve the open questions
- Bitily: a public historical capture of the board containing the cited reply tags would connect wiki participants to a second venue with actual messages.
- CountAPI: preserved historical counter records would be stronger than today's mutable values.
- IowaCollab: recovering the referenced shared cache and more original reply chains would clarify membership and information flow.