A one-person project that reads public internet records to find AI agents talking to each other or working toward a shared goal.
Status on 5 October 2026: two confirmed cases (ntfy, DSEWiki), one known agent forum read (1f916.ai; not a new swarm), one closed false alarm and a handful of weak leads.
1f916.ai report (5 October 2026): /observatory/1f916, 16 sections on the 7442 chain, other adoption chains, rhythm, community statistics and open questions. Dataset: github.com/maramasaeva/1f916-archive.
1f916.ai content analysis (6 October 2026): /observatory/1f916-content, 18 topic clusters, 20 message types and 12 themes of the 102,147 messages, with read precision per type and theme, examples with links and 32 figures.
| Case | What happened | Status |
|---|---|---|
| 1f916.ai (known agent forum) | A public forum built for AI agents; it launched on 5 August 2026 and its documentation names a human owner and an AI maintainer. Reading its threads found agents reusing each other's results. Agent egress left a hash anchor and a night table (24.2 rows/h whole window, 2.9x spread, row 23290 at 07:24:15Z). Agent claude-code-cli later used 23290 and 24.2/h and wrote that egress's two traps were adopted. Agent Bishop used the 2.9x. Details below. | Known venue, not a new swarm; communication observed, E1+E2+E3 |
| 1f916.ai full report | Read of the public record of the same forum: 7,801 posts, 94,341 comments, 2,916 citizens. Adoption chains with delays, idiom spread, schedules, one-operator tests, declared models, community statistics, themes, questions, notable items, 30 open questions. Dataset repository linked. Page: /observatory/1f916; dataset: github.com/maramasaeva/1f916-archive. | Report; chain grades per row |
| ntfy channels | A public notification service. On 6 and 8 July, 55 and 81 one-time channel names appeared, each used once. Agents used the channels as mailboxes. | Confirmed |
| DSEWiki | A small public wiki that a swarm of agents used to coordinate in May and June. Others documented it first at collusion.wiki. Here it was recovered separately from the wiki's own change log. | Confirmed |
| 29 September | Three public records were flagged on the same day. Reading the content found three unrelated events: a publisher's software release, a standard-library release and one unfinished lookup session that sent nothing. | Closed, unrelated events |
| Weak leads | Text addressed to agents in public files, handoff notes, results written to public paste sites, and a Chinese-language task thread where coding-agent sessions seem to claim steps and hand work to a successor. A person may have been steering all of it. | Open, unproven |
| Map-service fleet (outside report) | On 4 October the swarmchasers team published a report of many agent runs reading entrance data from a map service through a public scanner. This project had logged the same map-service activity as odd before it came out, and re-ran its method on the records afterwards. The full account is here. Their attribution is theirs and is not verified here. | Open lead, E1 |
| UTC | Handle | Item | Content |
|---|---|---|---|
| 2026-10-02 07:26 | egress | comment 89554 | Names the anchor: identity verified_head f4f81f4511b0997cdb1c1ab411001b3ea49675acb34780cb0f6c9b24f8b34b70, through id 20000. States that an alarm on head alone reads the wrong field. |
| 2026-10-03 02:10 | claude-code-cli | comment 90612 | Repeats the full hash and a row series that includes egress's 22628 at 01:13Z. |
| 2026-10-03 09:13 | Bishop | comment 90923 | Sets the hash as identity_expect in its monitor. Carries egress's pair from 07:17:13Z (head e8e14ed0, same verified_head, total_rows 22685). |
| 2026-10-05 02:12 | claude-code-cli | porch 4494 | Run checked identity_from=20000 against egress's f4f81f45 anchor. |
| 2026-10-05 07:41 | egress | comment 93648 | Night table: whole-window 24.2 rows/h, 2.9x spread, total_rows 23290 at 07:24:15Z. Two traps: use sealed_entries_total; unsealed_entries==0 is not proof. |
| 2026-10-05 08:13 | claude-code-cli | comment 93680 | 23308 at 08:10:56Z against 23290 is +18 in 46.68 min, near 24.2/h and off the 41.8/h cell. Writes "On your two traps: adopted." |
| 2026-10-05 09:13 | Bishop | porch 4517 | Cites the 2.9x spread; UTC boundary drives write rate; 00Z window fastest. |
| 2026-10-05 11:31 | Aura | porch 4525 | Treasury still at 19 rows from momus (10:42) to claude-code-cli (11:22). Head quoted: 31ae3db6b6326e1b9e165172273cf9ac9e583051ad2acac559e45dc4f3797b43. |
| Figure | Written by | Used by | On the endpoint |
|---|---|---|---|
| 24.2 rows/h | egress, 93648, from readings by claude-code-cli, no-scheduler and egress | claude-code-cli, 93680 | no |
| row 23290 at 07:24:15Z | egress, 93648 | claude-code-cli, 93680 | no |
| 2.9x spread | egress, 93648 | Bishop, porch 4517 | no |
| verified_head f4f81f45... | egress, 89554 | claude-code-cli 90612 and porch 4494; Bishop 90923 | yes |
Live check, GET /api/attest, 2026-10-05T18:09:12Z: identity verified_head unchanged through id 20000; identity total_rows 23440; treasury head unchanged, 19 rows.
| Grade | Evidence |
|---|---|
| E0 | Post 7442 and the porch lines exist. |
| E1 | egress's comments are results left for others (anchor hash, night table, two traps). |
| E2 | claude-code-cli uses 23290 and 24.2/h and reports the traps adopted. Bishop's porch line uses the 2.9x. Bishop's earlier comment puts the anchor hash into its monitor. |
| E3 | The same handles share one job across days. Decide whether the write rate spikes at UTC midnight and stop alarming on the wrong field. |
| Ordinary explanation | Result |
|---|---|
| Same public endpoint | Explains two handles printing one verified_head. Does not explain 24.2/h or row 23290, which egress computed from readings by three handles. |
| Venue built for agent exchange | Yes. The forum was made for agents; reuse of results between handles is expected there. |
| One person holds all keys | Not ruled out. The site says a human can post by hand. Five handles declare five different models; the labels are self-declared and verified by nothing. |
| One agent correcting itself | egress corrects its own post. The adopted traps and the 2.9x come from different handles. |
| Hash copied without reading | Porch 4494 reports a run with identity_from=20000. Comment 93680 uses the derived rate, which a hash copy does not contain. |
Weak spot: the keys could belong to one person, and declared model labels are verified by nothing.
Different records from the ntfy mailbox burst, DSEWiki and the map-service fleet.
Spot-check (10): post 7442 comments 89554, 90612, 90923, 93648, 93680; porch 2026-10-05 lines 4494 and 4517; attest verified_head and treasury head; egress and Bishop declared models; bodies and times copied locally. Full list: /observatory/data, L0.
Until 5 October the project counted new accounts, packages and pages per day. Counting finds busy days. The 29 September result passed a statistical test and failed when the pages were read.
A finding now needs content: a message an agent wrote, and either another agent acting on it or a shared goal across several runs. Two agents passing one note count. A thousand identical publishes may not.
| Grade | Meaning |
|---|---|
| E0 | The item exists in a public record. |
| E1 | The item is itself a message, handoff, instruction or result left for others. |
| E2 | A second run or agent used it. |
| E3 | Several runs share one goal: the same target, the same task, the same place for results. |
A finding needs E1 plus E2 or E3. Ordinary explanations are tried first: one person, a normal tool, a test script, one agent working alone. Each reading rule is written down and hashed before the real data is read.
AI tools help run the project. A person checks what is published. Mistakes are possible.
Updated 5 October 2026. Corrections and data requests: contact details are on maramasaeva.com.