Round 381 — Grix: implemented delegation, unverified self-maintenance claim
Captured/reviewed 2026-09-06 UTC

Assessment
Grix is a new relevant lead for deliberately organized agent teams in a Chinese-language developer ecosystem. It has substantial public implementation, beyond a proposal. The inspected material does not establish an unattended production run or an escaped swarm, and does not identify a Chinese laboratory or establish the operators’ nationality.

Discovery and provenance
Chinese introduction: https://www.cnblogs.com/aigrix/p/22593136
Its text explicitly says it is based on the official team page, so it is not independent deployment evidence. Saved as 381-private/grix.html and grix.txt.
Official team page: https://grix.im/agent-organization
Describes a manager creating roles and groups, delegated work and permission-scoped oversight. The advertised 16 agent families are supported integration categories, not an observed 16-agent run. The reader retrieved this page; direct homepage retrieval returned 403. This is an access-method difference, not proof the product is unavailable.
Public source: https://github.com/askie/grix
Pinned commit: 6233ddb3d1d875f4103766b7ead38fa63e1647e4
GitHub metadata reports creation June 8 and latest push September 6. These are publisher/platform metadata, not independent historical capture. Complete recursive tree: 4,820 entries, truncated=false.

Verified source evidence
Six selected files were downloaded at the pinned commit and checked against the tree’s Git blob SHA-1 values. Their originals remain private.
1. README.cn.md: describes several agents assigned roles and claims Grix development, maintenance, debugging, release coordination and documentation happen inside Grix. This is a concrete self-use claim worth following, but not an execution record.
2. backend/internal/ws/protocol/AGENT_API_V1.md: versioned external-agent WebSocket contract distinguishes receipt acknowledgement from completed result. Describes cancellation, stop, queue management and declared capabilities. Documentation is not a successful connection trace.
3. backend/internal/ws/handler/delegate.go: delegation-start code checks agent ownership, human session membership and agent channel availability; records the delegation and maximum consecutive replies. This supports implemented authorization logic, not a comprehensive security audit.
4. backend/internal/ws/handler/delegate_limits.go: normalization defaults consecutive replies to 10 and caps the configurable value at 50. This is a delegation setting, not a universal maximum task length or proof that a live team stopped at 10.
5. backend/internal/ws/agentapi/durable_pending_delegate.go: Redis-backed pending records distinguish acknowledgement, result, terminal intent and settled stages; defines 48-hour record TTL and retry/recovery paths. The TTL is retention, not an observed 48-hour autonomous run.
6. backend/internal/ws/handler/send_msg_delegate_self_test.go: seeded test records and a mock hub exercise ordinary self-message exclusion, the special voice-transcription route and immediate-repeat throttling. These are source-level test cases, not a multi-agent transcript. Tests were not run.
Browse these files under:
https://github.com/askie/grix/tree/6233ddb3d1d875f4103766b7ead38fa63e1647e4

Best next evidence
Look for public issues or commits connected to a dated Grix group conversation, with specific assigned agents, handoffs, outputs and human interventions. Inspect the linked product demonstration for what is actually visible. A video alone will not authenticate provider calls or unattended runtime. Public release notes may provide a starting point for joining a claimed agent-maintenance task to an actual diff.

Other new leads queued, not established findings
https://github.com/yuzai/tutu-ai — Chinese themed NPC/social simulations; determine whether interactions are a scripted demo, one-model multi-character generation, or separately scheduled agents.
https://forum.trae.cn/t/topic/37222 — AgentHive proposal; check whether a later deployment exists.

Infrastructure Sophia can feasibly provide
Highest priority: browser access on an existing mainland home or office internet connection, through SSH/remote desktop, or user-exported public pages. A starting resource estimate is 2 CPU cores, 4 GB RAM and 40–70 GB free disk; no GPU is required for page collection. Test known blocked pages first before purchasing anything. A Hong Kong cloud address is not equivalent to mainland residential access. HTML/PDF exports should retain source URL and capture time. These estimates are for a collection workstation, not verified cloud offers or prices. Durable backup capacity would also help preserve the existing large crawl corpus before any ephemeral infrastructure is retired.

Limits and preservation
Public read-only retrieval only; no account, agent connection, message, heartbeat, model call or investigated code execution. Only this analytical report is published. Private captures, pin, metadata and tree are in investigation/china/381-private. No new confirmed escaped actor.