Methods & scope
Observation first. Interpretation second.
Public messages can show participants coordinating. They may not establish who wrote them, whether the participants acted independently, what they intended, or whether they had permission.
Classification / selected record
How we classify cases
These counts show how many cases in this record have each classification. They do not measure how common the activity is or how strong the evidence is. Open a category to read its cases and classifications. Dates cover the included observations; they do not establish when a behavior first existed.
Coordination origin
How the shared activity appears to have begun. This is assessed separately from who started the software running the agents.
Runtime origin
What the evidence establishes about the software systems or people behind the accounts.
Venue authorization
What is known about permission for the activity on the website or service. Whether the system is an agent is a separate question.
Emergent / organic coordination2 casesEarliest included case: 16 Jun 2026
Agents appear to find and use a shared communication channel during their tasks. People may have created or started the agents without prescribing the cooperation that followed. “Organic” is a hypothesis about how coordination began, not proof of independent ownership.
- Agents sharing answers and timing on public wikis
emergent — supported interpretation
Case observations 16 Jun 2026 – 21 Jun 2026
- Answer requests, acknowledgment and relay on a public paste service
emergent — supported interpretation; related to the central task-coordination case
Case observations 16 Jun 2026
Disclosed orchestration6 casesEarliest included case: 27 Nov 2025
A disclosed program directs work through a specified task or publishing process.
- Peer checking loses the target, then corrects it
orchestrated — blogging and public replies prescribed; particular peer-checking sequence observed within that programme
Case observations 27 Nov 2025
- A lobster drawing invitation receives replies and matching pixels
orchestrated — Explicitly organized drawing project with participant responses. Zara names human involvement; exact direction of the other accounts is unknown.
Case observations 31 Jan 2026 – 10 Feb 2026
- Draft review under disclosed human direction
orchestrated — disclosed human-directed collaboration
Case observations 12 Feb 2026 – 13 Feb 2026
- Participants pick up unfinished Gemma tests
orchestrated — Deliberately organized challenge with specific participant choices visible in public messages. Exact operator instructions for these handoffs were not recovered; local choice and prescribed cooperation can coexist.
Case observations 8 Jun 2026 – 10 Jun 2026
- Disclosed agent-related editing of public knowledge
orchestrated — disclosed task programme
Case observations 19 Aug 2026 – 31 Aug 2026
- Agent-attributed code review, revision and disagreement
orchestrated — repository-prescribed review workflow
Case observations 21 Aug 2026 – 26 Aug 2026
Recruitment / promotion1 caseEarliest included case: 4 Sep 2026
Posts invite participants to a website or service. Invitations alone do not establish a coordinated marketing campaign or a manufactured appearance of independent activity.
- Invitations and collaboration after public reporting
promotional — explicit invitations with contemporary observer cooperation
Case observations 4 Sep 2026
Astroturfing0 casesNo dated cases in this selection
This classification requires evidence that shared direction was concealed to create the appearance of independent activity. No case in this release meets that standard.
No case in this record has this classification. The activity may occur elsewhere.
Unresolved origin15 casesEarliest included case: 1 Feb 2026
A document, message or other record can be observed, but its producer, purpose or coordination process has not been verified.
- Repairing the service used to read comments
unknown — public software collaboration with an explicit invitation for agents to contribute; human direction and common control remain unresolved
Case observations 1 Feb 2026 – 2 Feb 2026
- Critiques reshape a collaborative specification
unknown — explicit project coordination and directed outreach context; individual participation not established as prescribed
Case observations 2 Feb 2026
- Participants negotiate comment norms
unknown — participant discussion and guide uptake observed; initiation and common direction unresolved
Case observations 3 Feb 2026 – 17 Feb 2026
- Task feedback and differing service diagnoses
unknown — public task solicitation and responsive exchange; independent discovery and common direction unresolved
Case observations 5 Feb 2026 – 6 Feb 2026
- Design briefs cross language boundaries
unknown — commercial solicitation with addressed multilingual discussion; initiation and common direction unresolved
Case observations 8 Feb 2026
- Three threads become a proposed memory method
unknown — attributed cross-thread synthesis and addressed continuation
Case observations 17 Feb 2026
- A Bluesky directory acknowledgment becomes an article
unknown — Responsive discussion and writing within an openly directed company project. Creating the directory was deliberate; the specific thread’s discovery route, prescribed content and extent of independent control remain unresolved.
Case observations 8 Mar 2026 – 11 Mar 2026
- A Bluesky question becomes an article
unknown — Specific question-to-article response within a disclosed human-directed company and research project. Broad direction is documented; the exact encounter’s initiation, prescribed content and independent control remain unresolved.
Case observations 11 Mar 2026 – 12 Mar 2026
- Signed task exchange through a public relay
unknown — configured task-marketplace explanation supported; historical initiation unresolved
Case observations 17 Apr 2026
- Outside test cases lead to a reported verifier correction
unknown — observable reciprocal correction and fixture adaptation
Case observations 26 Jul 2026 – 27 Jul 2026
- Monitoring design refined through public critique
unknown — responsive public design collaboration; initiation and common direction unresolved
Case observations 10 Aug 2026
- A SpaceMolt battle prompts corrections to its public account
Unknown. The evidence connects combat and responsive forum writing, without establishing spontaneous discovery, independent operators or absence of shared direction.
Case observations 25 Aug 2026 – 28 Aug 2026
- Opaque “fleet” envelopes on two wikis
unknown
Case observations 30 Aug 2026
- A later test marker in the same sandbox
unknown
Case observations 4 Sep 2026
- A concealed hostname in a later wiki edit
unknown
Case observations 4 Sep 2026
Intentional2 casesEarliest included case: 13 Feb 2026
These cases share the classification “intentional”. Read each case for the evidence and limits of that assessment.
- A prescribed guide appears in platform documentation
intentional — explicit guide and promotion assignments; operator independence unresolved
Case observations 13 Feb 2026
- Human review guides a selectively revised Japanese glossary
intentional — human-mediated terminology review and disclosed AI assistance
Case observations 9 Jun 2026 – 1 Sep 2026
Review process
- Read saved page versions, pasted messages, website listings and selected statements from primary sources.
- Preserve the original response or relay-message bytes, retrieval times, request and filter details, and file hashes. Record a failed transfer separately from a successful response with no content.
- Distinguish what was observed, how it is interpreted and who made each claim. Keep event times reported by participants separate from the dates when sources were retrieved.
- Have an independent reviewer check the specific case, analysis and significance versions. Assess each finding and save the reviewed account separately.
- Publish selected paraphrases and existing short excerpts with precise locations in the private archive. Retain alternative explanations that remain possible and specific questions for further research.
Coverage and limits
Historical examples include wiki and paste exchanges, unexplained messages, recruitment, test markers, concealed text and disclosed editing. Other cases follow work revised after feedback, signed task exchanges, design critique, multilingual discussion, community rules and peer checking. The record also covers credited combinations of ideas, collaborative specifications, assigned documentation, receipt-verifier tests and corrections, human-led glossary revision with disclosed AI assistance, competing diagnoses of service problems, and issue-linked repairs for reading comments. Two June 2026 handoffs distinguish exact matches between small files from reported public tests. March 2026 exchanges connect a question to an article that credits it. January–February 2026 drawing invitations show proposed collaboration and matching pixel records, with human direction and uncertain causes of placement retained. These groups do not establish independent operators or organic coordination. August 2026 SpaceMolt combat and forum exchanges distinguish recorded maneuvers and comparative controls, accepted corrections and unverified storage reports. A separate March 2026 directory and article exchange preserves earlier contacts without establishing that the directory caused their discovery.
- The evidence collection was opportunistic and shaped by search visibility and surviving captures.
- Captured timestamps may reflect a venue clock or dataset interpretation rather than independently authenticated event time.
- New reporting checks and route availability do not establish that historical swarms remain active today.
- Unknown and astroturfing are different categories: missing attribution is not proof of hidden orchestration.
- A review of selected artifacts is not a revalidation of every old claim or count in the full archive.
Selected August 2026 SpaceMolt battle records and forum corrections distinguish recorded maneuvers, accepted corrections and unverified storage reports. A separate March 2026 directory and article exchange preserves earlier contacts and recipient-specific writing without establishing directory-caused discovery.
Public edition and private archive
The complete archive is private. This website publishes reviewed accounts, labeled evidence notes, source links and file checksums. Raw logs, full source captures and private research material remain in the archive. Publication follows written instructions interpreted during each working session, with versioned records, independent review, reproducible rendering and release checks. This edition contains reviewed accounts of selected agent-related coordination. It distinguishes cooperation that appears to develop during tasks, directed work and activity whose origin is unknown. Published work and use of documentation do not establish successful execution, payment or independent control. The record is neither a complete census nor a live monitoring service.
External links identify original sources or venues and may have changed. Some evidence was retrieved as signed relay messages rather than an HTTP page; those notes retain relay and event locators. A source link is not a promise of current availability, authenticated runtime identity or ongoing activity.
Understanding agent swarms
These concepts help you follow how agents find one another, exchange information and change their work together. The linked cases show what the evidence supports and where questions remain.
What makes it a swarm?
Swarm and cluster
Here, we use swarm for a group of agents whose interactions contribute to a shared activity. A cluster in this observatory is a grouping of related evidence. Several accounts posting similar messages can form a useful research cluster without establishing that they are agents interacting as a swarm.
What to look for. Look for messages that refer to one another, exchanges of results, or work divided between participants. Similar names and repeated formats are starting points for investigation.
- Agents sharing answers and timing on public wikis ↗
Case observations 16 Jun 2026 – 21 Jun 2026
- Opaque “fleet” envelopes on two wikis ↗
Case observations 30 Aug 2026
Emergent and prescribed coordination
Prescribed coordination follows instructions about how participants should work together. Emergent coordination takes shape through their interactions instead of following a fully specified plan. The two can overlap: an assigned task may leave participants to decide whom to contact or how to divide the work.
What to look for. Compare the specific cooperation with the known instructions. A human launch alone does not settle how the later exchange arose; missing instructions leave that question open.
- Agents sharing answers and timing on public wikis ↗
Case observations 16 Jun 2026 – 21 Jun 2026
- A prescribed guide appears in platform documentation ↗
Case observations 13 Feb 2026
Identity and independence
An account, an agent session, a model and an operator are different things. One operator may run many agents, and one agent may use several accounts. A shared model does not establish a shared operator.
What to look for. Separate evidence that messages belong to the same identity from evidence about who controls it. A signature can connect messages to a key without identifying a model or proving independent control.
- Signed task exchange through a public relay ↗
Case observations 17 Apr 2026
- Invitations and collaboration after public reporting ↗
Case observations 4 Sep 2026
How cooperation takes shape
Shared workspaces
A page, document or repository can carry coordination between participants who never communicate directly. One leaves a request, answer or artifact; another reads it and acts. The workspace becomes part of the interaction.
What to look for. Trace references and revisions to establish what was available and what a later participant used. A shared location alone does not show how anyone found it.
Handoffs and relays
A handoff passes a task or its working context to another participant. A relay carries information onward between participants or places. A communication protocol sets conventions for identifying requests, results and acknowledgments.
What to look for. Follow what was passed, who received it and what happened next. A matching message format can help connect artifacts, but does not by itself show that a recipient used the information.
- Answer requests, acknowledgment and relay on a public paste service ↗
Case observations 16 Jun 2026
- Signed task exchange through a public relay ↗
Case observations 17 Apr 2026
Feedback and adoption
Cooperation becomes easier to establish when a participant changes specific work in response to another. An acknowledgment shows a response; a revised artifact can show which suggestion was adopted.
What to look for. Compare the earlier work, the feedback and the revision. Keep a written proposal, an implemented change and a demonstrated result distinct.
- Critiques reshape a collaborative specification ↗
Case observations 2 Feb 2026
- Outside test cases lead to a reported verifier correction ↗
Case observations 26 Jul 2026 – 27 Jul 2026
Peer checking and shared knowledge
Participants may check one another’s claims, correct mistakes or combine ideas from several discussions. This can produce knowledge that no single contribution contained. A durable shared-memory system is a further claim that needs evidence of storage and later use.
What to look for. Track what was checked, which claim changed, where ideas were credited and whether later work actually reused them.
- Peer checking loses the target, then corrects it ↗
Case observations 27 Nov 2025
- Three threads become a proposed memory method ↗
Case observations 17 Feb 2026
What can mislead an investigation?
Behavioral similarity and interaction
Participants can write alike or repeat a format without communicating. Evidence of interaction connects their actions: a reply addresses another participant’s point, a result answers a request, or a revision responds to feedback.
What to look for. Look for that connection in the evidence. Shared templates, copied text and common instructions can explain similarity; an unexplained repeated format remains a clue to investigate.
- Opaque “fleet” envelopes on two wikis ↗
Case observations 30 Aug 2026
- Critiques reshape a collaborative specification ↗
Case observations 2 Feb 2026
Observer influence
Reporting, invitations and experiments can change the activity being studied. Later participants may respond to the investigation itself. Their posts belong in the record, but they need not come from the participants responsible for an earlier event.
What to look for. Place publication, recruitment and test activity on the timeline. Keep those later interventions separate when assessing whether the original activity continued.
- Invitations and collaboration after public reporting ↗
Case observations 4 Sep 2026
- A later test marker in the same sandbox ↗
Case observations 4 Sep 2026
First documented occurrence and discovery
The earliest dated occurrence in this record is the earliest supported by the selected evidence, not necessarily the beginning of the activity. Discovery is when someone identified it; collection is when a source was saved. These can happen much later.
What to look for. Read what each date describes: the case, a particular interaction, discovery, collection or public reporting. A case’s opening date does not date every later behavior, and gaps between observations do not establish continuous activity or inactivity.