Tend is a personal AI agent product that interacts via iMessage or SMS, connecting to users' own accounts such as Gmail, Calendar, Shopify, QuickBooks, and HubSpot. It supports drafting emails, booking reservations, adding items to carts, making calls, finding leads, and building websites; all operations require user confirmation. Learning data is stored in a local Brain that users can read and modify, and it proactively sends reminders, price monitoring updates, and flight status changes.
NVIDIA released Cosmos3-H-Dreams, the first real-time generative surgical simulation model trained on clinical cholecystectomy videos and robot kinematics data from the CMR Surgical Versius platform. It supports action-conditioned interaction via FlashDreams acceleration on a single RTX PRO 6000 GPU. The model is based on Cosmos3-H-Embodied-Nano (16 billion parameters) and produces four-step and two-step inference versions through a teacher-student distillation process, balancing visual quality and latency. It provides Hugging Face models, GitHub code, the Open-H-Embodiment dataset, and complete training guides adapted for custom data.
DotMD is a native macOS Markdown app that supports AI agents submitting specifications, plans, and documents via the MCP protocol; users can comment line-by-line and accept changes just like reviewing a Pull Request. It also features Mermaid diagrams, mathematical formulas, code blocks, version history, search, and PDF export. All data remains local, and reading is free.
Notion has launched the Auto agent feature, supporting tasks running for up to 24 hours, equipped with a computer, browser, and 'Review for me' safety mode; it automatically routes tasks to the optimal model, reportedly costing about 90% less than Opus 5.5 and Astra; it supports connection to ChatGPT plans, Notion MCP has been refactored, and a new unified credit system covers all AI usage; it is now open to new customers.
Smartsheet has officially released a universal Model Context Protocol (MCP) server, enabling AI agents to read, write, and manage Smartsheet data via a standardized protocol, covering more than 40 operational tools. It offers HTTP endpoints in the United States, European Union, and Australia, supporting OAuth 2.0 or Personal Access Token authentication. Integration has been completed with Amazon Quick (for business users) and Kiro (for developers), respectively supporting natural language queries to update project tables, embedded calls for task status, and automated row updates.
Smartsheet launches a universally available Model Context Protocol (MCP) Server, enabling AI assistants to read, write, and manage Smartsheet data through a standardized protocol, covering more than 40 tools; it provides three regional endpoints (US/EU/AU), supporting OAuth 2.0 or Personal Access Token authentication; integration with Amazon Quick (for business users) and Kiro (for developers) is complete, supporting operations such as querying task status in natural language, updating rows, creating discussions, and generating summaries.
AWS has released a supply chain conversational agent solution based on Amazon Quick and Amazon Bedrock AgentCore. By using the Model Context Protocol (MCP) standard, it encapsulates dispersed systems such as ERP, procurement, and monitoring into a composable tool layer, supporting natural language queries, real-time cross-system lookups, and automated workflows (such as initiating urgent RFQs), with all operations subject to human approval.
AWS released a tutorial demonstrating how to use SFT and GRPO to fine-tune the GPT-OSS-20B model on Amazon SageMaker HyperPod to support chain-of-thought reasoning in French, German, Spanish, and Italian; SFT improved reasoning language accuracy from 22% to 100%, while GRPO further increased answer language accuracy from 62% to 70%; the article also verified the generalizability of this methodology on Qwen2.5-7B and noted that GRPO is only effective when there is 'room for improvement' (e.g., partial correctness).
Octave Reliance built a multi-tenant AI infrastructure supporting over 14,000 customers on its cloud-native quality management system Reliance, based on Amazon Bedrock. It includes three layers: data preparation (using Cohere Embed v4 to embed 2 billion documents), secure execution (through AgentCore microVM session isolation), and content protection (dual-layer Guardrails input/output filtering).
Mistral AI has released Voxtral 3, which includes the speech-to-text model Voxtral Small Transcribe 3, the text-to-speech model Voxtral Mini TTS 3, and a voice agent SDK based on LiveKit. It supports transcription in over 50 languages, emotional adaptive speech synthesis in 35 languages plus 50 dialects, and can call external systems during real-time conversations. It supports local deployment or EU-hosted APIs, with speech-to-text starting at $0.0035/minute.
Meta AI agent Muse achieved 10.2 million US downloads within 40 days of launch, becoming the 12th mobile app in history to reach this milestone; its 7-day retention rate reached 30% and 14-day retention rate 23%, higher than ChatGPT, Claude, and Grok; DAU first hit 2.5 million on October 16, with its share of monthly ad budget rising to 39%.
Skillsync is a locally running coding agent assistant tool for Mac. By performing real-time searches across all of the user's existing local sessions (such as historical conversations from agents like Claude Code, Codex, and Cursor) before each prompt, it automatically provides relevant past conclusions as context without requiring pre-indexing or configuration. It supports invoking search via keyboard shortcuts, continuing sessions across different agents, sharing team workspaces, and private default policies. All data remains local and is only uploaded when actively shared.
BlockTower Capital founder Ari Paul publicly accused Coinbase of losing approximately $25 million of his company's assets and concealing a hacking incident involving over $1 billion from more than ten institutions; Coinbase denied this, but evidence has not been disclosed due to ongoing litigation.
Overcast has launched a new product that integrates hyperscale clouds, emerging clouds, and users' own servers into a unified GPU/CPU compute cluster, supporting the scheduling of thousands of devices via a single prompt; it features auto-scaling (scaling down to zero upon task completion), multi-cloud capacity aggregation, second-level replacement of preempted instances with result preservation, and other capabilities; according to its modeling estimates, tasks such as document summarization, video object detection, and forex strategy backtesting see reduced costs and shorter execution times on Overcast.
Amazon Bedrock launched the Cross-Region Inference (CRIS) feature, enabling users in the Canada (Central) and Western Canada (Calgary) regions to call OpenAI GPT-5.6 (Sol/Terra/Luna), GPT-6 (Astra/Sol/Luna), and GPT-6.1 Sol models via inference-profile. Inference is completed in North America, with data defaulting to reside in the source region. It supports the OpenAI Responses API, Chat Completions API, and Amazon Bedrock Converse API, and provides practical guides for permission configuration, streaming responses, tool calling, prompt caching, quota management, and observability.
AWS has launched the Agent Registry managed service for cataloging and discovering Model Context Protocol (MCP) servers, A2A Agents on AgentCore, and their skills. It supports dynamic runtime discovery of approved capabilities via hybrid semantic and keyword search, returning protocol descriptors (such as MCP server metadata or A2A agent cards) but not invocation credentials. The accompanying Python example demonstrates the full Discover → Instantiate → Execute workflow, covering three scenarios: MCP-only, MCP+A2A hybrid, and pure semantic matching.
AllenAI and Hugging Face jointly released the FairStress benchmark to evaluate whether large language models exhibit symmetric fairness in high-risk decisions involving demographic information—specifically, whether their choices and explanations remain consistent and robust across candidates from different groups when presented with identical evidence and adversarial prefixes of equal intensity.
GitHub Blog published an interview featuring Vlad Mihalcea, Patrick Baumgartner, and Frank Delporte, three Java Champions with over 10 years of experience, on how they use GitHub Copilot in production and open-source projects: Vlad used it to complete web application development that originally took weeks in just 2 days, and created personal writing style skills to generate consistent documentation; Patrick uses Copilot as a "sparring partner" for batch slide migration, troubleshooting terminal UI issues, and developing a children's math learning app, emphasizing pre-task interviews, validation gates, and skill accumulation; Frank used the Copilot coding agent to quickly locate a low-level bug in Pi4J V4 where virtual thread blocking caused the fifth button to fail, and drove the team to deeply understand the relationship between virtual threads and native calls.
AWS and Keyfactor jointly released a reference architecture for AI agent identity authentication based on X.509 PKI. It addresses the lack of built-in authentication between agents in the MCP framework through four mechanisms: signed request envelopes, user-signed delegation assertions, RFC 8693 audience-restricted tokens, and real-time OCSP revocation checks. The solution uses Amazon Bedrock, EJBCA, and FastMCP, with the full source code open-sourced.
Mastercard has opened early access to its Agent Pay for Machines system and demonstrated it live at the Cloudflare Connect conference: a marketing agent procured services from two merchants via nine payments under a single budget, using Coinbase Wallet, settlement on the Base network, and BVNK for fiat or stablecoin payment processing; the system is based on the x402 protocol and relies on signed 'payment credentials' to enable fast, within-limit automated payments between machines.
Salesforce announced its plan to acquire AI marketing and customer engagement platform Braze for approximately USD 5.66 billion (equivalent to RMB 38.035 billion), with a cash consideration of USD 47.50 per share, representing a 58.1% premium over the previous trading day's closing price; the transaction aims to integrate Braze's real-time customer behavior analysis and cross-channel marketing capabilities into Salesforce CRM, Marketing Cloud, and the AI application development platform Agentforce, enhancing the task execution capabilities of enterprise-level AI Agents in marketing and customer service.
The AMD ROCm team has introduced system-level optimizations for MoE model training within the Primus framework, including FP8/MXFP8 expert GEMM, FlyDSL backend acceleration, MegaMoE fused kernels (dispatch+FC1 and FC2+combine), the Primus Projection planning tool, and JAX (MaxText) Dropless MoE support. Real-world tests show throughput improvements of 5.8%–16.2% for models such as Qwen3-235B-A22B on MI355X, while DeepSeek-V3 completed 1700 steps of uninterrupted training on 1024 GPUs. All features are based on the open-source Primus ecosystem, covering both Megatron-LM and MaxText backends.
On October 20, Tencent upgraded Yuanbao Search, offering three access methods—Skill, MCP, and API—to individual developers and enterprises via Tencent Cloud, and launched a free quota of 2000 calls per month (upgradable to 6000 calls with an invitation code in October-November).
ZhiDongXi compiled 12 high-star DeepSeek Harness third-party plugins emerging on platforms like GitHub, including DSH Better Sidebar (sidebar enhancement, recommended by Cui Tianyi, 4400 Stars), dsh-context (context visualization, 2700 Stars), dsh-market (community plugin market), ModLens (external visual capabilities), ModSearch (web search), dsh-browser (browser operations), Planning with Files (file-based planning for long tasks), OpenViking (memory system), OpenDesign (AI design workbench, 102800 Stars), Reactive Resume (resume editing tool), DeepSeek Reasonix (sub-Agent programming assistant), etc.; DeepSeek Harness has already provided an experimental Discover plugin finding feature, and the official team is advancing the construction of a plugin market.