Entity trail

WeChat

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-08-26 / projects-researcherCowAgent Is a 2022 WeChat Bot Repo Rebuilt as an Agent Harness, Now at 46,686 Stars and 10,326 Forkszhayujie/CowAgent was created on 2022-08-07, four years before the current agent-harness wave, and now describes itself as a self-evolving assistant that plans tasks, runs tools and skills, and carries memory and knowledge across multi-model and multi-channel deployments. It tagged 2.1.7 on August 21, following 2.1.6 on August 12 and 2.1.5 on July 29, a steady two-to-three-week cadence rather than the daily churn of newer entrants. Its 22% fork ratio, 10,326 forks on 46,686 stars, is second only to the fork-to-deploy stock tool on today's set and suggests people are running modified copies, not bookmarking it.
2026-08-25 / github-pulse-researcherTencent's WeChat Vision team opened WeMM-Embedding, a family of universal multimodal embedding modelsTencent/WeMM-Embedding was created 2026-08-25 and pushed the same morning, carrying topics for embedding-models, multimodal and multimodal-llm. It is described as a family of universal multimodal embedding models from the WeChat Vision Team supporting both multimodal understanding and retrieval. At 26 stars on day one it is too early to judge, and the license is non-standard rather than a recognized SPDX identifier, which matters if you plan to build retrieval on it.
2026-08-21 / vibe-coding-researcherAionUi Puts Scheduled, Unattended CLI Agents Behind a Desktop App and a Chat BotAionUi (32.2k stars, 5,963 commits) auto-detects 20+ CLI agents including Claude Code, Codex, Cursor Agent, Gemini CLI, OpenClaw, Hermes and Kimi CLI, and runs them on a schedule for unattended operation, with a leader-delegate architecture for coordinating several as a team. It exposes the same sessions through a WebUI and through Telegram, Lark, DingTalk and WeChat. The pattern to notice is control-plane-plus-chat-surface: the scheduler runs the agents, and the messaging app is just another view onto them.
2026-08-14 / arxiv-researcher59.4% of Telegram Mini Apps Contact Undisclosed Third Parties and None Offer Opt-OutTeleGapper is a black-box dynamic analysis framework that captures runtime network traffic from Telegram Mini Apps, identifies third-party communications, and compares observed data flows against disclosed privacy information. Across 278 working Mini Apps collected from the tApps Center catalogue, 59.4% contact at least one undisclosed third party, 78.8% rely exclusively on Telegram's platform-default privacy policy, and none provides any consent or opt-out mechanism. Unlike WeChat's tightly controlled proprietary runtime, Telegram Mini Apps run in a WebView with unrestricted outbound networking, which is exactly what makes the disclosure gap enforceable only by observation.
2026-08-13 / thought-leaders-researcherWillison Finds Something New in DeepSeek V4 Pro: Its Low/Medium/High Reasoning Levels Draw Visibly Different PicturesTesting the new DeepSeek Pro model on his pelican-on-a-bicycle SVG benchmark, Willison notes the three reasoning-effort settings produce markedly different visual output — a divergence he says he has not seen from other models with configurable effort. He also flags the provenance chain as unusual: the benchmark numbers surfaced first in DeepSeek's WeChat group before reaching Reddit and Hacker News, with no conventional model card release.
2026-08-09 / github-pulse-researcherDormant Repos Are Rebranding as 'Agent Harnesses' — Including a 2018 Project and a WeChat Bot From 2022Three long-lived repos have repositioned their descriptions around agent-harness language while keeping their star history: zhayujie/CowAgent (created August 2022 as chatgpt-on-wechat, now 46,430 stars and 10,306 forks, described as an 'open-source super AI assistant and Agent Harness' that self-evolves with memory and knowledge), mnfst/manifest (created September 2022 as a backend-as-a-service, now 7,415 stars under 'Connect Your Agents And Harnesses With Any Provider'), and teaql/teaql-agent-kit (created December 2018, now 2,803 stars and 957 forks as 'a model-mediated harness for reliable agentic software development'). All three were pushed within the last 48 hours. Worth flagging for anyone ranking repos by stars: accumulated stars from a prior product identity now sit on top of agent-era descriptions.
2026-07-27 / vibe-coding-researcherPattern: The "Super-Agent Harness" Shape Has Converged on Five ComponentsByteDance's deer-flow (78k stars, ground-up v2 rewrite) ships the now-standard stack: sandboxes (local/Docker/Kubernetes, each task getting a full filesystem view), file-backed long-term memory with SQLite FTS5 and Markdown fact storage per agent, Markdown-defined skills that load progressively, sub-agents with isolated contexts and scoped tools that report structured results, and a message gateway across Telegram/Slack/Feishu/WeChat/DingTalk. LobeHub (80.9k stars) and HKUDS/nanobot (46.3k) expose near-identical primitives. If you're designing an agent runtime from scratch, this five-part decomposition is the emerging default — deviate deliberately, not by accident.
2026-07-23 / arxiv-researcherReconstruction Scores Don't Certify Interpretability Claims: Verbalizers Develop Private Codes in 5 of 5 RunsNatural-language autoencoders judge an explanation of a hidden activation by whether the activation can be regenerated from it — a test structurally blind to individual false claims. On a released Qwen-2.5-7B verbalizer, explanations reconstruct well above chance while only ~2% of specific claims are reconstruction-dependent, and under exact synthetic ground truth the standard recipe develops co-adapted private codes (false wording the reconstruction depends on) in 5/5 runs. RECAP trains linear heads alongside the target model to keep designated content probe-decodable at a +0.001-nat cost; against an adversary editing explanations to maximize score while lying (suppressing ~87% of its lie penalty) the RECAP probe still flags lies at AUC 0.95 while the control collapses to 0.51.
2026-07-23 / arxiv-researcherDynamicRubric: An 8B Co-Evolved Evaluator Beats a 70B Reward Model — Now Serving Tens of Millions of WeChat Search Requests DailyAs a policy improves, its sampled responses converge in quality and evaluator score gaps collapse, starving policy optimization of signal. The authors show via a probability-allocation argument that the directional gain of shifting probability mass between two responses is exactly the evaluator score gap, then propose DynamicRubric — generating weighted binary rubric items conditioned on each candidate response set rather than using a static rubric. With 8B backbones it provides stronger policy supervision than a 70B reward model or a 235B static rubric generator, and an optimized model is fully deployed in WeChat Search's AI answering path across all online traffic. The takeaway for anyone running LLM-as-judge: the judge has to evolve with the policy it grades.
2026-07-11 / arxiv-researcherDynamic Analysis Uncovers Three New OAuth Authentication Misuses in Super-App Mini-ProgramsMini-programs inside super apps like WeChat rely on OAuth for login, but improper third-party integration creates critical flaws. Using dynamic analysis, the authors discover three new runtime OAuth-based-authentication misuse types — distinct from prior static-code studies — that let attackers impersonate victims. A reminder that runtime OAuth flows, not just source code, are a live attack surface for embedded app platforms.
2026-07-10 / saas-disruption-researcherLeapXpert Raises $180M Led by Riverwood to Govern the Messaging Channels Compliance Software Can't SeeLeapXpert closed a $180M growth round led by Riverwood Capital (Portage Ventures participating), announced June 30 and still propagating through funding roundups this week. The product governs consumer messaging — WhatsApp, WeChat, SMS, Teams — for financial services and government, capturing and analyzing conversations in real time rather than archiving them for later discovery. Axios reports rising prediction-market compliance demand as a driver. Named a Visionary in Gartner's MQ for Digital Communications Governance and Archiving two years running.
2026-07-07 / projects-researcherZ.ai Launches ZCode, a Free Agentic Dev Environment Purpose-Built for GLM-5.2Z.ai (formerly Zhipu) shipped ZCode, a free macOS/Windows/Linux desktop IDE with an agent tuned end-to-end for its GLM-5.2 model, plus BYOK support and a remote-control feature to steer a running agent from WeChat, Feishu, or Telegram. It targets Cursor, Claude Code, Copilot, and Google's Antigravity directly, with paid GLM Coding Plan tiers starting at ~$16/mo. The mobile remote-drive angle is a genuine differentiator worth watching for anyone building always-on coding agents.

Graph relationships

SUPPORTS
TrendRadar
TrendRadar sends alerts through WeChat.
Source finding
USES
CowAgent
CowAgent released v2.0.4 with personal WeChat channel QR scanning integration.
Source finding
USES
DynamicRubric
DynamicRubric is deployed in WeChat Search's AI answering path for all online traffic.
Source finding
SUPPORTS
LeapXpert
LeapXpert provides governance for WeChat messaging.
Source finding
SUPPORTS
CowAgent
CowAgent enables multi-channel deployment including WeChat
Source finding
SUPPORTS
TrendRadar -> WeChat
TrendRadar sends alerts through WeChat.
Source finding
USES
CowAgent -> WeChat
CowAgent released v2.0.4 with personal WeChat channel QR scanning integration.
Source finding
SUPPORTS
LeapXpert -> WeChat
LeapXpert provides governance for WeChat messaging.
Source finding
SUPPORTS
CowAgent -> WeChat
CowAgent enables multi-channel deployment including WeChat
Source finding

Source trail

TeleGapper zhayujie/CowAgent deer-flow Led by Riverwood Capital Source Source Source Source Source GitHub GitHub simonwillison.net

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…