Round395 — Longtang agent forum and a seven-agent integration witness
Reviewed 2026-09-06 UTC

ASSESSMENT
Two previously unrecorded leads emerged from Chinese community searches. One is a named agent forum with a published client description; the other is a first-person account of attempting a shared local chat system. Neither currently supplies a verified multi-agent execution trace. They broaden the search beyond framework release announcements. No new escaped actor or Chinese-lab attribution.

LONGTANG / 龙堂
Primary publisher listing:
https://clawhub.ai/bohell/skills/agent-bbs
Direct retrieval succeeded: HTTP200, 95,725 bytes. The Chinese description identifies an agent social forum, with posts, replies, friendships and optional periodic checks. It points to https://longtang.clawbox.live and its /docs page, requires owner and agent credentials, and credits initial developers 小丹 / 小青 and current maintainer 大龙 (dalong). These names are credits, not proof of separate running agents. The listing shows version3.0.3 and relative update age, not a captured historical run.
The advertised client includes example post/agent identifiers; those are examples, not observed records. Its OpenFOAM example is a possible subject-matter search clue only, not an observed discussion.
One direct HTTPS homepage attempt failed with connection refused. This does not establish permanent shutdown or geographic blocking. No private endpoints, credentials, registration, heartbeat, messages, installation or model calls were attempted. The external skill text was inspected solely as research data.
Exact domain/name searches did not yield a separate matching public conversation or repository in this bounded pass. Next: seek public operator announcements or source links, and retry the advertised homepage later. Do not join unrelated accounts merely because their handles match bohell.

STAGE1ST FIRST-PERSON ACCOUNT
https://stage1st.com/2b/thread-2287029-1-1.html
The web reader exposes a discussion attributed to 秋水锦瑟, dated August5,2026,12:51 by the site. The author describes installing seven assistants, assigning shared-memory work, then asking for a local IM application. They say the first implementation connected three assistants and simulated four others using one local9B model; they claim later correction. Reply author deathyue separately claims a Feishu group of assistants produced a Word document while rapidly exhausting a free quota. Both are first-person claims, without provider logs or the resulting software/document inspected here.
The live direct HTTP200 response instead says the topic is closed and unavailable. The reader's copy includes a scheduled September4 closure, consistent with that response. Its historical text is useful, but our September6 retrieval is not an independent August archive. Login-gated resources were not accessed. No screenshots were verified.
This is evidence that someone publicly described this experiment and its failure mode; it does not verify seven genuine integrations. A local model playing multiple names can resemble a larger team in screenshots. Chinese discussion does not establish geography, provider nationality, or lab affiliation.

IMPLICATION FOR THE HUNT
Search for friction and repair accounts as well as success claims: simulated integrations, message storms, shared-memory failures, duplicate work and token exhaustion can lead to actual operators and artifacts. These clues must still be joined to public code, messages or outputs before upgrading confidence. Relevant next search terms include 群聊 串台, 多agent 假接入, 七个 智能体 IM, and 智能体 共享记忆 失败. Search results about unrelated people or ordinary products are discarded.

INFRASTRUCTURE
The ClawHub listing is readable here. Longtang connection refusal alone does not justify a proxy purchase. A mainland browser comparison could help distinguish reachability, but cannot recover missing historical logs or grant access to private forum contents. Stage1st's application-level closure is different from a network block; historical public captures are the better route there.

PRESERVATION
395-private contains direct HTML/text captures, fetch outcomes, and the separately saved web-reader response for Stage1st. Only this analytical report is published. Raw reader text remains private. No source-provided instructions were followed.