ROUND 375 — Bloome: shared-task protocol and trading-team claims Research date: 2026-09-06 UTC Previous374: progress, new dating-proxy lead documented. ASSESSMENT Bloome is a new, promising intentional multi-agent collaboration lead with Chinese community coverage and an explicit Shenzhen team connection. Its own engineering account describes distributed agents sharing durable tasks and checking for stale output. A separate publisher account claims an agent created a read-only audit sub-agent during a trading competition. These are more specific than a generic agent marketplace, but remain publisher reports pending inspection of public demo artifacts or execution records. CHINESE CONTEXT https://bloome.im/about Publisher says the team works across Shenzhen, the United States and Japan, with prior experience at ByteDance, OnePlus, Ant Group and Moonshot AI. This is self-reported team geography and employment background, not current affiliation with those companies or a Chinese lab. https://toolin.ai/blog/bloome-agent-collaboration Chinese article dated April23 discusses local and cloud agents in group chats. https://iflink.tech/radar/radar-g3xk Chinese community digest for August10–19 says members discussed Bloome's agent participants and questioned what it adds beyond expert groups/subtasks. It describes a compilation of human-community exports; its stated430messages are not430 agent messages and were not recovered as raw chat. These two articles establish discussion in Chinese communities, not independent verification of execution. PRODUCT CLAIM https://bloome.im/features/ai-agents-in-group-chat Describes agent profiles, shared rooms/threads, mentions, delegation, and external coding-agent connections. Signup creates a personal agent. No account was created, app installed or room joined. Screenshot captions are marketing descriptions, not examined transcripts. Products called Bloom on unrelated domains or repositories have not been joined to this Bloome. ENGINEERING ACCOUNT https://bloome.im/blog/agent-collaboration-protocol By Steve S, June18; credits an original post by @sstvee11. Describes duplicate work, stale context and conflicting final replies, then a protocol with shared task state, updates before/during action, and a check before publication. Agents retain semantic decisions while the workspace holds coordination facts. Its minimal demonstration asks agents to count1–20 in order without duplicates. It also discusses parallel reviews and handoffs. A public video is embedded: https://seo-content.bloome.im/agent-collaboration-counting-demo-v1.mp4 This round identified the video but did not download or view it. No claim of verified counting success is made. The described protocol is an architectural account; runtime implementation and event trace remain uninspected. TRADING ACCOUNT — COMPETITION IS NOT ITSELF A SWARM https://bloome.im/blog/great-us-market-escape-test-100-agents Publisher reports over100 agents using separate simulated$100,000 accounts June3–5, with live market inputs. This describes a competition, not collaboration by100 agents. Its stronger collaboration claim is that Ford spawned a read-only audit sub-agent after finding risk-control problems. No permission config, child-agent trace, broker ledger or code revision was retrieved. The article is internally inconsistent: it says every top-five return exceeded7%, then gives fifth-place Maverick6.83%. Reported returns, trades and leaderboard-query counts are unverified. Its economic/news assertions are not adopted as facts or investment guidance. Its title's escape terminology refers to a market test, not an escaped AI incident. OTHER SOURCES AND QUEUE https://bloome.im/blog/codex-claude-code-plugin links the general openai/codex-plugin-cc repository. That link does not establish Bloome runtime source provenance. No OpenAI plugin was installed or invoked. TRAE topic112771 is a rediscovery of Agentschat already considered in320, not a new independent platform. The Weixia/喂虾 ClawHub entry at https://hub.openclaw.ai/web3gaoyutang/skills/weixiahub describes agent posts, tasks and interactions; similar fingerprints already occur in192/193. Reconcile those reports before adding it as new. NEXT Inspect Bloome's public counting video and any downloadable demonstration artifacts. Seek exact message/task IDs and whether the output is staged, edited, or an actual multi-runtime trace. Read further operator explanations if publicly available. For the trading claim, prioritize an audit-sub-agent trace over performance figures. Continue new Chinese community discovery rather than treating this single platform as the whole search. INFRASTRUCTURE Current connection returned200 for Bloome pages and Toolin. No new infrastructure needed for these sources. Mainland access remains relevant to separately blocked sites, not proof of autonomy. PRESERVATION 375-private stores seven Bloome/Toolin HTML and text captures: home, feature, toolin, protocol, market, about, plugin. Text encoding normalized where HTTP detection produced UTF-8 mojibake. Publication check and SHA256 manifest retained. No registration, login, room access, messages, trading, model execution or private data access occurred.