2026.08 · Social AI Radar

QQ Bot official plugin adds an interactive question channel with buttons for group and direct chats

Tencent tencent-connect team merged an interactive question channel into the official DeepSeek Harness plugin for QQ Bot on August 24. When an AI agent requires user input, it can invoke ask_user_question in QQ group chats and direct messages to deliver questions with Inline Keyboard buttons or text options, intercepting answers in middleware to advance a multi-question state machine with timeout cleanup and avoid queue deadlocks.

Why it matters: This equips AI agents in group chats with proactive questioning and structured guidance capabilities. Interactive buttons lower user effort for giving input, establishing a foundation for multi-agent workflows and multi-step human-in-the-loop collaboration.

QQ Bot · China · First seen 2026.08.25 10:45 Beijing time

Updated from 2026-08-18-qq-bot-rich-media-vision

Teams launches GitHub Copilot cloud agents in public preview for shared coding in chats and meetings

GitHub announced the public preview of GitHub Copilot in Microsoft Teams on August 21. Users can mention @GitHub in Teams channels, group chats, meetings, and direct messages to start a cloud agent session, allowing all participants to ask questions, add context, and direct edits, while members with write permissions can trigger code changes, create pull requests, and execute tasks asynchronously in secure cloud sandboxes.

Why it matters: This embeds asynchronous AI coding agents directly into daily team discussions and meetings, enabling non-technical stakeholders and developers to jointly define requirements, review changes, and steer development in the same chat context without switching between specialized dev tools and messaging platforms.

Microsoft Teams · Global · First seen 2026.08.25 10:45 Beijing time

ChatGPT Work and Codex add Apple Messages search, drafting and approved sending

OpenAI released the Apple Messages plugin in the ChatGPT desktop app for macOS on August 20. It can find chats, summarize messages, prepare replies and send through Messages on the Mac. By default, both the message and its recipients require user approval before sending. The plugin is available on all plans, but works in ChatGPT Work and Codex rather than regular ChatGPT chats.

Why it matters: This moves AI beyond drafting into reading relationship context and completing the final step before a message is sent, while keeping recipient and content approval as the default boundary. Because Apple Messages spans iMessage, SMS and RCS, it is a high-sensitivity agent entry point into mainstream personal and team communication records.

ChatGPT Work / Codex · Global · First seen 2026.08.24 09:11 Beijing time

Slack Code launches shared code channels where teams and AI agents build together

Slack launched Slack Code on August 20. When a user mentions a supported coding agent in a channel or direct message, the agent can create a temporary code channel for that task. Teammates can continue prompting in the same space, share files, inspect code diffs and live previews, comment on individual lines and approve work before it ships. The feature is available on all Slack plans with a supported agent installed, initially supporting Claude, Devin, GitHub Copilot and Vercel, while OpenAI integration is listed as coming soon.

Why it matters: Slack turns agent work that would normally stay in a private window into a multiplayer process that teammates can see, redirect and review. Temporary channels preserve project context and an audit trail while placing code, documents and previews beside the team conversation, making this a concrete native workspace for human-agent collaboration.

Slack · Global · First seen 2026.08.24 09:11 Beijing time

Meta AI connects to Instagram and Facebook business data for content insights and reports

Meta began rolling out new small-business features on August 19. Users can opt to connect Meta AI directly to professional Instagram and Facebook accounts, Meta ad campaigns and Google Workspace. The assistant can analyze engagement metrics such as reach, saves, shares, comments and profile visits, compare public content from similar brands, and turn the findings into presentations, documents or spreadsheets. The capabilities are available free across meta.ai and the Meta AI mobile and desktop apps, with support for recurring analysis tasks and reminders.

Why it matters: The update gives a general-purpose AI assistant first-party interaction context from social accounts instead of limiting it to generic advice based on public posts. For creators and small businesses, content performance, audience feedback, competitor observation and team reporting become one conversational workflow, while the social platform's proprietary data becomes part of the AI product's differentiation.

Meta AI · Global · First seen 2026.08.24 10:34 Beijing time

Xiaohongshu deploys OneModel to unify feed, advertising and merchant ranking

A technical paper published on August 19 says Xiaohongshu has deployed OneModel in its production ranking system. It combines user behavior across organic recommendations, advertising and merchant services into long sequences, learns shared user and content representations with a long-context backbone, and then modulates ranking objectives by scenario. Online A/B tests report a 0.33% increase in time spent and a 1.25% engagement lift in Explore Feed, an 8.18% advertising CTR lift, and a 1.1867% increase in daily gross merchandise value for merchant recommendations.

Why it matters: Xiaohongshu is no longer keeping user understanding isolated across community content, advertising and commerce, instead sharing the full behavioral trajectory through one ranking foundation. For a content community, that means follows, likes, comments, viewing and purchasing signals can influence distribution more deeply together, while raising the governance challenge of competing objectives affecting one another.

Xiaohongshu · China · First seen 2026.08.24 09:11 Beijing time

Kuaishou reports 797.3 million Q2 MAU as AI agents move deeper into recommendations and live streaming

Kuaishou's second-quarter 2026 results, released on August 19, show average DAU of 412.3 million for the quarter ended June 30, compared with 408.9 million a year earlier, and average MAU of 797.3 million, compared with 714.8 million. The company also said the Kuaishou App is using AI-powered placement to improve acquisition and retention, and introduced AgentX, a self-evolving agent for recommendation systems. Its live-streaming experience further upgraded AI Interaction Assistants, digital avatars, content understanding and real-time multimodal gift recommendations. The number of mutual followers using private messages grew by more than 15 percent year over year, while users sent more than 6 million AI gifts during the quarter.

Why it matters: This disclosure moves AI beyond the standalone Kling product and into the relationship, distribution and live interaction systems serving nearly 800 million monthly Kuaishou users. Recommendation agents, mutual-follower messaging and live AI tools now appear in the same operating picture, showing how a content community can use AI to reshape traffic allocation, creator services and paid interaction at the same time.

Kuaishou · China · First seen 2026.08.20 10:15 Beijing time

QQ Bot's official DSH plugin adds rich-media handling and an optional vision tool

Tencent's tencent-connect team merged rich-media understanding into the main branch of the official DeepSeek Harness plugin for QQ Bot on August 18. The new code can download images, videos and files from direct chats, group chats and quoted messages, expose local paths to the agent, and adds a qqbot_describe_image tool that calls a configured multimodal model for image understanding. Media is cleaned up after 24 hours by default. The vision tool is disabled by default and the latest public npm package remains version 0.4.0 from August 16, so this item is recorded as code merged with the installable release still pending.

Why it matters: This extends AI agents in QQ groups from text-only memory to shared image, video and file context, allowing members to collaborate around the same visual material. Keeping the vision model off by default and expiring downloaded media also shows that multimodal agents in messaging systems must balance capability, cost and privacy boundaries.

QQ Bot · China · First seen 2026.08.24 09:11 Beijing time

Reddit tests AI-narrated video versions of posts and comment threads

Reddit began testing a new playback format for posts in selected English-language communities on August 17. Chosen text posts show Read and Play modes, with AI using different synthetic voices to narrate the original post and selected comments alongside captions and a visual template. The page is labeled 'Real conversation voiced by AI,' while the original post and comments remain unchanged. The experiment currently reaches only selected posts and users, and Reddit is manually choosing which content to convert.

Why it matters: Reddit is turning the community discussion itself into a platform-generated content format rather than merely summarizing it for search. Selecting comments, assigning voices and trimming context directly shape how community members are represented, creating early governance pressure around consent, accuracy and opt-out controls.

Reddit · Global · First seen 2026.08.24 09:11 Beijing time

WeCom's open source CLI reaches 1.0 for terminal-based AI agent workflows

The WeCom team released version 1.0 of its open source wecom-cli on August 17 and published @wecom/cli 1.0.0 on npm. The MIT-licensed tool lets people and AI agents operate WeCom from the terminal across documents, smart sheets, messaging, contacts, tasks, meetings and calendars, and provides an installable CLI Skill.

Why it matters: This extends WeCom's relationship and collaboration data beyond the app interface into a terminal entry point that agents can invoke. Packaging messaging, documents, calendars and contacts as commands and a Skill makes it easier to connect group communication and office execution in one automated workflow.

WeCom · China · First seen 2026.08.20 20:18 Beijing time

Lark aily rebrands to Doubao Work Partner with deep office suite integration

ByteDance officially rebranded Lark's Aily assistant to Doubao Work Partner on August 14. Following the organizational merger of Lark and Doubao product teams, Doubao Work Partner deeply integrates with Lark messaging, Base, docs, meetings, and calendars to execute cross-tool tasks, summarize group chats, suggest replies, and support custom multi-agent virtual teams for enterprise collaboration.

Why it matters: ByteDance is consolidating its AI brand under Doubao, marking Lark's transition from a collaboration tool to an agentic enterprise infrastructure. By connecting chat, docs, and structured tables, AI is moving from an external sidebar into the central workflow of enterprise operations.

Lark · China · First seen 2026.08.24 12:55 Beijing time

Updated from 2026-07-24-feishu-aily-team-agents

WeChat tests an AI Processing entry for images in direct and group chats

A sequence of interface screenshots published on August 14 shows a new AI Processing button when some WeChat users open an image in either a direct chat or a group chat. The entry offers image enhancement, information extraction, image editing and more, and can hand tasks to Xiaowei for text extraction, summaries, image translation, document generation, background replacement and object removal. The entry remains in limited testing, and WeChat has not announced a broad rollout.

Why it matters: Images are among the most common pieces of shared content in direct and group chats. Placing Xiaowei inside the image viewer means AI is no longer only a separate conversation destination. It can work on content already moving through a relationship network, combining understanding, editing and follow-on sharing within the same messaging flow.

WeChat · China · First seen 2026.08.20 10:15 Beijing time

Updated from 2026-07-23-wechat-xiaowei-beta

QQ Bot adds an official DeepSeek Harness plugin for direct and group chats

The official QQ Bot plugin package for DeepSeek Harness first became public on August 14, and Tencent QQ formally announced the integration on August 15. After setup, the bot can act as an AI assistant in direct messages and group chats. Each person and group has separate persistent conversation memory, sessions continue after a restart, and users can switch models without losing context, while initial setup supports QR code binding and group chats can require an @ mention before the bot replies.

Why it matters: This brings a general AI agent into the existing direct and group relationship network on QQ, so users do not need to open a separate app. Separate memory and mention-only replies turn context isolation and response boundaries into explicit group settings, showing how messaging platforms are becoming important entry points for agents.

QQ Bot · China · First seen 2026.08.17 15:59 Beijing time

WeChat details Xiaowei's WeLM models, with 80B total parameters and 3B active

WeChat's official account detailed the WeLM model family built by the Weixin team on August 12. WeLM-80B, which is already deployed in the native Xiaowei assistant, has 80 billion total parameters and activates 3 billion per inference. It supports chat and search, use of native WeChat functions and access to Mini Program services. WeLM-617B remains in development with 617 billion total parameters and 23 billion active, targeting more complex WeChat tasks such as intelligent Mini Program development and tool generation for Xiaowei. Xiaowei itself remains in a small-scale prototype test.

Why it matters: The parameter disclosure completes the link between Xiaowei's model scale and its product entry points. WeChat is not simply placing a general model inside a chat box. It is using sparse activation to control inference cost across a relationship and service network at billion-user scale, while preparing different model sizes for native functions, Mini Programs and tool generation.

WeChat · China · First seen 2026.08.20 10:15 Beijing time

Updated from 2026-08-12-wechat-xiaowei-welm-scale

Tencent discloses 1.439 billion WeChat MAU and the WeLM model behind Xiaowei

Tencent's second-quarter 2026 results, released on August 12, show that combined MAU for Weixin and WeChat reached 1.439 billion as of June 30, up 2% year over year and 0.5% quarter over quarter. Tencent also confirmed a small-scale prototype test of Xiaowei inside Weixin, powered by the Weixin-customized WeLM model with a focus on user privacy, Weixin use cases and inference efficiency. The earnings presentation says the agent can draw on the social graph, knowledge graph, merchant reach and payment functionality, with user intervention and multi-step confirmations as safeguards. The official materials do not disclose WeLM's parameter count.

Why it matters: The disclosure puts the real scale of WeChat's relationship network and the model boundaries of its embedded agent in the same set of official facts. Xiaowei is not merely adding general question answering. Its design centers on privacy, inference cost, the social graph and service transactions, showing how a super app can connect AI directly to relationship and commerce infrastructure.

WeChat · China · First seen 2026.08.13 09:26 Beijing time

Updated from 2026-07-23-wechat-xiaowei-beta

Circle AI opens beta access to all customers with shared team Projects

Circle said every customer would be able to opt in to the Eclipse beta by August 1, and its current pricing page lists Circle AI as a core feature beginning with the Professional plan. Circle AI Projects lets teams group chats, files, instructions and project memory in one workspace, invite teammates to share context and hand off conversations, and use the AI as a shared collaborator for community setup, operations and moderation.

Why it matters: This moves community operations AI from an administrator's personal assistant into a shared layer for team operations. Project context, teammate collaboration and the AI's understanding of community conversations now sit in one place, bringing AI directly into community organization and governance.

Circle AI · Global · First seen 2026.08.06 09:25 Beijing time

LinkedIn adds AI slop reporting and changes its writing assistance

LinkedIn product chief Hari Srinivasan announced a set of measures against low-quality AI content on July 31. Members can now report a post as seeming like AI slop, while LinkedIn is strengthening classifiers that reduce low-quality material in recommendations and out-of-network content. The company will also remove the Enhance your post feature from posts and messages in favor of proofreading that preserves the author's voice, and will test private analytics feedback when members see content as inauthentic or heavily AI-assisted.

Why it matters: AI content governance is expanding from platform-only detection into a loop that combines community feedback, distribution controls and private guidance for authors. It also treats AI use and low-quality content as different problems, aiming to limit mass-produced noise while preserving creator control over voice.

LinkedIn · Global · First seen 2026.08.03 09:45 Beijing time

Pally connects Teams and upgrades Slack and WhatsApp message triage

Across July 30 and 31, Pally shipped new cross-platform messaging capabilities. Users can connect Microsoft Teams through a secure browser, with one Microsoft sign-in covering both Outlook and Teams, while the intelligent triage used for scheduled email updates now also covers Slack threads and WhatsApp chats. Busy Slack channels are reviewed in batches, direct mentions still arrive immediately, bot-only channels can be demoted to mention-only alerts, and users can set plain-language approval rules across email, messaging and shopping.

Why it matters: Personal AI is moving from a single chat entry point into an attention and coordination layer across workplace channels and private messaging. Unifying triage, channel noise reduction and approval policies can reduce cross-platform misses while giving users clearer control over when an agent alerts them or acts on their behalf.

Pally · Global · First seen 2026.08.02 09:06 Beijing time

Teams adds AI follow-ups for frontline teams and cross-meeting recaps

Microsoft Teams published its July update on July 31. Follow Up in Frontline Agent can send one request to as many as 20 people, automatically collect responses by a deadline, consolidate the results and track non-responders. The Meeting Recaps app brings intelligent recaps from the past 30 days into one place and can generate a podcast-style audio summary across multiple meetings.

Why it matters: Team AI is moving beyond passively summarizing one meeting and into the loop of prompting, collecting and consolidating work from several people. A manager can turn one request into a trackable group workflow and recover shared context across meetings, reducing coordination loss across chats and calendars.

Microsoft Teams · Global · First seen 2026.08.01 09:20 Beijing time

ChatGPT opens five-person workspaces to academic research teams

On July 29, OpenAI opened applications for ChatGPT for Academic Researchers. Eligible faculty and postdoctoral researchers can receive 12 months of complimentary access to a dedicated ChatGPT workspace for as many as five verified members from the same institution. The workspace includes business data protections and ChatGPT Pro-level usage limits, and approved applicants can invite four collaborators while keeping team research in a separate space. The program does not include API credits.

Why it matters: Institutional verification, a dedicated workspace and small-team seats turn ChatGPT from an individual research assistant into a controlled collaboration environment. The product pairs shared AI context with a clear data boundary, reducing the friction of coordinating research through shared accounts or scattered personal chats.

ChatGPT · Global · First seen 2026.08.05 09:42 Beijing time

Threads adds Meta AI conversations to direct messages

Meta announced on July 28 that one-to-one conversations with Meta AI are now available in Threads direct messages. Users can search for @meta.ai to start a private chat or share Threads posts, images and videos into the conversation for follow-up questions. The exchange is visible only to the user and is currently available on mobile, not the web.

Why it matters: Threads is embedding an AI assistant in the path from public discussion to private understanding. People can continue asking about shared posts without leaving their social and content context, turning direct messages into an entry point for both sharing and AI-assisted interpretation.

Threads · Global · First seen 2026.08.03 09:45 Beijing time

WeChat tests a photo-to-Xiaowei entry inside Scan

A set of beta interface screenshots published on July 27 shows that some WeChat users with Xiaowei access can see a new Photo to AI entry at the bottom of Scan, marked with a NEW badge. After taking a photo, Xiaowei opens in a half-screen view so the user can keep asking about the image. Examples include translation, remembering the moment or a friend, finding similar products and drafting a Moments post. The entry remains in staged testing.

Why it matters: Google Lens lets people ask about what they see through a camera or image, ChatGPT accepts images as conversational inputs, and Doubao offers photo recognition and visual understanding. WeChat's distinct move is to place a similar snap-and-ask entry inside the high-frequency Scan flow, bringing visual questions into an everyday camera path in a messaging product. These products do not have identical capabilities or user journeys.

WeChat · China · First seen 2026.08.02 20:54 Beijing time

Updated from 2026-07-23-wechat-xiaowei-beta