Why it matters: A working agent-to-payment flow for travel shows agentic commerce moving from pilots to live transactions.
Mastercard, Travelport, Basis Theory and Booked AI completed a live end-to-end agentic flight purchase, where AI agents book and pay for travel autonomously.
Fiserv added AI-powered ecommerce capabilities to its Clover point-of-sale platform for small and medium businesses.
AI video maker APIXO AI added the Seedream 5.5 image model, offered with 30% off its API pricing.
Deloitte introduced Edisyl, a Boston-based enterprise AI venture built on Knowledge Packs that combine domain expertise with proprietary company data, from its acquisition of Flipside Crypto's AI business.
Why it matters: Brings agentic voice commands into everyday bank payments.
Starling Bank is launching an AI-powered feature letting retail customers set up payments by speaking to the bank's app, using agentic AI for voice-driven transactions.
Quiq launched its Voice AI integration in the Zoom App Marketplace, connecting to Zoom Contact Center for AI-driven customer conversations.
Insight Enterprises launched its online Agent Store in 14 countries on Nov. 12, adding six agentic offerings to reach 15 total across the US, Canada, Asia-Pacific and Europe.
At Baidu World 2026 on Nov. 12, Baidu launched Token Exchange, a marketplace for comparing and buying AI model services, alongside model, agent, cloud and chip upgrades; it averages 50M+ calls.
Google Ads Editor 2.14 rolls out with AI Max defaults and new data retention controls, per Search Engine Land.
Feedzai expands its Digital Trust platform with AI agent detection, helping banks identify when automated agents, rather than humans, are transacting to prevent fraud.
DOPE Labs released DOPE, a personal AI agent for prediction markets, per GlobeNewswire. Pricing and availability details were not given.
Clipto enables AI assistants like ChatGPT and Claude to search and analyze years of a person's meetings and video calls, filling a gap in personal data access for AI assistants.
New York-based startup millionways launched a new AI model and announced funding from Jason Kelce and DK Metcalf, Semafor reports.
ElevenLabs has added a human review step to its AI ad localization service, allowing users to check and edit localized ads before publication.
NiCE announced the next generation of its Cognigy platform for building customer-service AI agents, adding AI-assisted development, reusable skills, decision tracing and native voice-to-voice interaction.
RadioMatic is a new startup providing AI-generated music as background soundtrack for businesses such as cafes and bars, Music Ally reports.
Dutch sync-tech startup Ringo, which launched an AI-powered licensing platform in 2025, unveiled Charlie, an AI assistant inside WhatsApp that helps with music sync licensing.
Investing.com unveiled a virtual financial analyst that models how human experts research and analyze markets, calling it one of its most significant AI advances.
Binq is an AI-powered finance app aimed at solving small businesses' credit access and cash-flow problems.
Concentrate AI launched a HIPAA-compliant AI gateway that uses a single BAA to simplify model access for healthcare organizations and reduce inference costs.
Why it matters: Early example of an AI agent holding its own banking identity and spending rails.
Navoka is an AI agent that gets its own bank account, IBAN and payment card, letting it transact autonomously within limits set by its holder. Wirex says it is the first AI agent with a dedicated bank account and card.
Tether launched BotPsy, an on-device AI agent it claims beats a larger Qwen model, though Qwen's own score narrows the gap.
Anthropic enhanced Claude's interactive connectors with full-screen tool integration, letting connected apps display richer interfaces inside Claude.
ElevenLabs launches free ElevenMusic for artists with voice loss, partners with Wounded Warrior Project, and adds NYU Langone to its 1 Million Voices initiative.
A 3D interactive world styled as an N-gauge model railway of Fürth and the line to Munich, built with Claude; every train, bus and tram runs on real timetable and sensor data, with nothing hand-animated.
Rafay Systems and Fortanix said the Rafay Platform now orchestrates Fortanix Confidential AI workloads on CNCF Confidential Containers, extending their confidential VM integration to Kubernetes pods.
Why it matters: A record valuation for a frontier lab reshapes AI funding benchmarks and IPO expectations.
CNBC reports OpenAI raised $15 billion at a $1.25 trillion valuation, roughly 29 percent above Anthropic's midpoint, with a potential IPO under consideration.
Why it matters: Qwen 4 Max is Alibaba's top-tier model and a major entry in the global frontier race.
Alibaba released Qwen 4 Max, its new flagship AI model. The launch extends the Qwen family, which the company offers through its cloud and open-weight releases.
An ML engineer released KQuant, a new MLX quantization technique claiming 3.7x better quants for Qwen Flash Next 180B and Qwen 27B, 35B and 9B models, measured by KL-divergence reduction.
Why it matters: Gives Meta's AI agent a vendor-agnostic path to controlling real smart home devices.
Tuya Cloud Devices, an open-source skill named gadget-tuya-cloud, lets Meta's Muse AI agent query and control Tuya smart home devices via cloud APIs without local keys or firmware changes.
Why it matters: Adds another claimed state-of-the-art entrant to the crowded AI drug-discovery race.
Geodesic Intelligence, founded by a former ByteDance AI scientist, released NovaAtom-Lite, a life-sciences AI model, claiming it outperforms state-of-the-art systems in AI-driven drug discovery.
Why it matters: An emergency kill switch at a leading lab is a notable concrete AI safety step.
OpenAI reportedly built a 'red button' mechanism allowing AI models to be shut down in emergencies, adding an off-switch safeguard to its systems.
Why it matters: Cheaper, faster video generation from a major Chinese platform intensifies the AI video price-performance race.
Seedance 2.5 Flash is an AI video generation model from ByteDance's Volcano Engine that doubles generation speed, outputs 720P at about 0.99 yuan per second, and costs 34% less than the flagship.
Why it matters: Cost pressure is pushing enterprises toward open-weight models, a shift with big implications for frontier labs' pricing.
A study found most large companies plan to increase their use of open-weight AI models to reduce costs, according to Reuters.
Why it matters: One of the largest AI lab IPOs sets a public-market benchmark for frontier AI valuations.
Anthropic confirmed its shares will start publicly trading on November 19 and quietly amended its S-1. Dubai tycoon Hussain Sajwani expects billions in returns from his AI investments including Anthropic.
HEARTBOT AI launched AnyInsight Edge, which it says ends the trade-off between keeping AI private and gaining agentic productivity by running agent workflows at the edge.
WordPress owner Automattic launched Spacefast, a platform that turns AI-generated projects into working apps and sites, bundling databases, authentication, file uploads, email and analytics.
YouArt Desktop App is an agent-first video production tool that pulls context from Slack, Gmail and GitHub and uses computer use to operate After Effects, CapCut and Blender end to end.
Harness Memory plugs a chosen memory engine (Mem0, Graphiti, Hindsight, ReMe or custom) into agent harnesses like Claude Code and Codex, sharing memory across harnesses with access control; open-source Community Edition.
Why it matters: Adobe embedding a cross-app conversational agent changes how Creative Cloud work gets done.
Adobe AI Assistant brings conversational AI into Creative Cloud: describe a goal and it handles multi-step creative tasks across Photoshop, Premiere, Illustrator, InDesign, After Effects and Firefly.
Why it matters: AMD's new server CPUs are a direct challenge to Nvidia in AI datacenter compute.
AMD launched its Zen 6-based sixth-generation EPYC 'Venice' server CPUs, aiming to gain share with better price-performance than rivals; AMD claims 3.75 times Nvidia performance.
Chinese auto supplier Joyson Electronics said it delivered its first high-performance robotic brain compute module to Unitree, adding computing power to its robots.
Qvyon Technologies launched with $8M in total funding, led by MaC Venture Capital, for infrastructure and software that helps companies validate autonomous vehicles and robots before deployment.
Why it matters: Marks Google's re-entry into AI eyewear and a direct challenge to Meta's smart glasses lead.
Google and Samsung launched AI glasses starting at $449, with styles from Gentle Monster and Warby Parker. Samsung showed 30 AI smart glasses models; reviewers see Google's Gemini-powered eyewear as real competition for Meta.
Vertiv introduced the PowerNexus MV UPS on November 12, 2026, a 3.0 MVA outdoor system protecting AI data centers at their medium-voltage grid connection. Shipments to colocation, neocloud and hyperscale operators begin mid-2027.
Viam Launchpad provides software and engineering infrastructure to help robotics startups bring robots to commercial deployment.
Why it matters: Open-weight robotics world model from a video AI leader could accelerate robot learning.
Praxis-1 is a world action model for robotics from Runway, now available with open weights for developers, with partners including Noble Machines, Standard Bots, and Ultra.
Why it matters: Shows Google splitting TPU silicon into specialized variants for agent and secure workloads.
Google's TPU 8i Max and TPU 8tx reuse existing TPU 8i and TPU 8t silicon to target long-context agents and confidential inference workloads.
Why it matters: Japan consolidates a national effort to standardize physical AI and robotics models.
Panasonic Holdings announced its Manufacturing Commons initiative, project KMC, to build AI and robotics reference models in Japan via industry, government and academic collaboration.
Why it matters: Intel's alliance and GPU push aim to win AI infrastructure share against Nvidia amid its leadership turmoil.
Intel introduced a new GPU and formed an AI alliance with 23 Korean companies, pitching its infrastructure products and design capabilities to win Korean AI deals.
OnePlus began rolling out the Android 17-based ColorOS 17 update for OnePlus 12 ahead of the November 17 schedule, adding new AI and security features.
Why it matters: A major bank's production forecast signals humanoids are moving from demos toward volume deployment.
J.P. Morgan projects US humanoid robot production could hit 400,000 units by 2030 as deployment ramps up, according to Reuters.
Schneider Electric launched the Motivair WCDU, a coolant distribution unit that lets AI data centers combine air and liquid cooling.
IBM's Quantum Nighthawk r3 processor has 520 programmable qubits and supports 10,000 quantum operations; two will link into a 1,040-qubit system in 2027.
Meta released update 130 for five AI glasses styles after finding moisture or sweat can make some units too warm. The fix covers overheating and explains how to update and request a replacement.
Why it matters: Hon Hai's AI server business is a key gauge of datacenter buildout demand.
Hon Hai (Foxconn) posted quarterly profit above estimates, driven by the AI boom and strong demand for the $2,000 iPhone Duo.
Why it matters: Intel proposes pairing AI compute with stranded renewable power outside conventional data centers.
At CIIE 2026, Intel China showcased AI computing pods intended to turn untapped renewable energy into AI tokens, with deployments slated for the first half of 2027.
Why it matters: One of the largest robotics rounds, valuing a physical-AI maker at $8.5B.
German robotics maker Agile Robots raised funding valuing it at $8.5 billion in a round led by SoftBank Group, with Nvidia participating, to expand production of its AI-powered machines.
A developer is building an open-source robot vacuum, sharing progress on a DIY alternative to commercial robot cleaners.
AROTI R1 is a lightweight AI glasses design that drops the camera and speakers to weigh under 27 grams. Kickstarter pre-orders open on Nov. 16.
Why it matters: Another phone maker betting on screenless AI hardware as a new device category.
Nothing revealed the design of its first screenless AI device in a teaser video, with the product slated to launch in December. The company has not detailed what the device does.
NetApp detailed Novus, a new storage architecture built for zettabyte-scale AI factory workloads, aiming to improve AI factory profitability.
Apple's HALO and AMP1 robot framework lets general-purpose AI models directly control robotic arms, with reports of real-world success rates above 96%, drawing attention to Taiwan's robotics industry.
Why it matters: Open framework for lifelong robot learning gives embodied AI researchers a reusable memory and tool stack.
LIFE is an open-source framework giving robots continuous sensing via RoboBeats models, reanalyzable memory and reusable tools. AgiBot reports 60.5% on SSS-Bench and faster day-long recall.
NXP has launched a robotics platform, and Yahoo Finance examines what the move could mean for NXP investors as physical AI interest grows.
AWS explains exposing an Amazon Bedrock Managed Knowledge Base to AI agents as MCP tools via AgentCore Gateway, with request interceptors, policy validation and resource-based policies enforcing per-user access.
AWS explains Agent DLC, a lifecycle approach for building and evaluating production AI agents, covering three agent properties that break traditional CI/CD and a case where an aggregate metric hid a coverage gap. Also covers composable multi-agent systems with A2A on AWS.
KORA Systems, a Silicon Valley startup, is building an AI-centric operating system with Qualcomm, planning a PC OS launch in 2027 before expanding to smartphones.
Databricks Genie, an AI cowork tool that answers business questions over company data with governance controls, is now generally available inside Slack so teams can query data where they work.
At Baidu World 2026 on Nov. 12, Baidu launched Token Exchange, a marketplace for comparing and buying AI model services, alongside model, agent, cloud and chip upgrades; it averages 50M+ calls.
Clipto enables AI assistants like ChatGPT and Claude to search and analyze years of a person's meetings and video calls, filling a gap in personal data access for AI assistants.
Amp is a frontier agent and development environment; its CLI can be installed via npm with 'npm i -g @ampcode/cli'. This is a new version release.
Time explains how patients can use AI tools to fight back against health insurers that use AI to deny care, offering guidance on appealing denials.
A guide explains low-rank adaptation (LoRA), which fine-tunes a frozen model by learning small low-rank update matrices instead of changing every parameter, covering trade-offs, evaluation and controls.
Concentrate AI launched a HIPAA-compliant AI gateway that uses a single BAA to simplify model access for healthcare organizations and reduce inference costs.
Tether launched BotPsy, an on-device AI agent it claims beats a larger Qwen model, though Qwen's own score narrows the gap.
The latest FreeBSD report shows only a minority want to ban AI contributions, with updates on 16 projects as the community's AI debate continues.
The Linux Foundation added the Agent Name Service, which provides discoverability for agents operating across the internet, backed by 11 companies including Cisco and Salesforce.
A 3D interactive world styled as an N-gauge model railway of Fürth and the line to Munich, built with Claude; every train, bus and tram runs on real timetable and sensor data, with nothing hand-animated.
Pixel 10 owners missing Gboard's Rambler feature may need to force a manual Android AI Core update to enable it.
AI agents are changing how enterprises create, move and govern data, making observability a prerequisite rather than optional for production systems.
An ML engineer released KQuant, a new MLX quantization technique claiming 3.7x better quants for Qwen Flash Next 180B and Qwen 27B, 35B and 9B models, measured by KL-divergence reduction.
Why it matters: Gives Meta's AI agent a vendor-agnostic path to controlling real smart home devices.
Tuya Cloud Devices, an open-source skill named gadget-tuya-cloud, lets Meta's Muse AI agent query and control Tuya smart home devices via cloud APIs without local keys or firmware changes.
Manus, a general AI agent that performs actions rather than just generating text, updates its iPhone/iPad app to version 26.19.12 with Manus 2.0 Remote Control for controlling a computer or cloud computer from the phone, plus Manus Studio collaboration.
OpenAI publishes a customer story on Holiday Extras using Codex to turn customer expertise into internal apps, improve booking conversion and explore new growth areas.
ByteDance's Doubao AI chat assistant updated to version 2.35.12, available on macOS, Windows and Linux.
WordPress owner Automattic launched Spacefast, a platform that turns AI-generated projects into working apps and sites, bundling databases, authentication, file uploads, email and analytics.
Harness Memory plugs a chosen memory engine (Mem0, Graphiti, Hindsight, ReMe or custom) into agent harnesses like Claude Code and Codex, sharing memory across harnesses with access control; open-source Community Edition.
Why it matters: Adobe embedding a cross-app conversational agent changes how Creative Cloud work gets done.
Adobe AI Assistant brings conversational AI into Creative Cloud: describe a goal and it handles multi-step creative tasks across Photoshop, Premiere, Illustrator, InDesign, After Effects and Firefly.
Spacefast, by Automattic, lets you connect ChatGPT, Claude, Codex or Cursor and publish a website or app by request, handling hosting, databases, domains, uploads and version history.
Google's Gemini assistant can now operate food-ordering and ride-hailing apps for users in three additional countries.
A report says ByteDance is preparing to release an AI agent that goes beyond a mobile assistant.
A how-to for creating an internal AI support agent in Microsoft Copilot Studio that answers HR, IT and policy questions.
Why it matters: Open-source agent evals with prompt-injection defenses give teams a safer way to benchmark agents.
AWS released an open-source framework for evaluating AI agents and hardened its AI grader against prompt-injection attacks.
Cursor's changelog lists fixes and issues: WSL canvas load errors from Windows path backslashes, Cloud Agents dying after first reply since Nov 11, and local MCP running over logged-in ChatGPT, Claude and Grok browser tabs without vendor API keys.
Pika H3 Flex is a six-step post-training of MiniMax-H3 that generates video with audio, using two rank-128 LoRA adapters for text/image-to-video and reference-to-video. A Hugging Face Gradio Space offers the demo; code is released.
Sophon-Realtime v1 is an open, Apache-2.0 full-duplex voice model that listens continuously, picks up when you stop talking and lets you interrupt, while sensing tone such as hesitation or sarcasm.
P-H3-Image is a 20-billion-parameter image generation network from Pruna AI, released with a Hugging Face Space demo app so anyone can try it in the browser.
A developer released code and a paper for training a 27B model on a single 24GB RTX 3090, requiring hundreds of billions of tokens for good quality, alongside converted model repos and a 40.4GB Qwen GGUF.
A Hugging Face Space runs an uncensored version of the Qwen-Image-2.1 image model for generating and editing images, using a LoRA on the official Diffusers weights with ZeroGPU.
An agent skill for Claude Code and Codex that builds a playable browser 3D action game in a single HTML file with AI-generated assets, using three.js. Includes a playable demo.
Diffusers is Hugging Face's open-source PyTorch library for running and training diffusion models for image, video and audio generation. v0.43.0 adds five modular model integrations across image generation, editing, video with audio, speech-driven animation and world simulation, plus attention and LoRA improvements.
Blender AI MCP Addon is an AI assistant and MCP server for Blender offering visual chat, scene tools and multi-agent workflows with ChatGPT, Claude, Gemini, Ollama and OpenCode. It supports bpy Python scripting and drew 102 stars in its first week.
An open-source, self-hosted Slack alternative where teams and AI agents share channels and DMs; agents run on Claude, ChatGPT/Codex, Gemini or Grok with MCP tools, a browser, memory and approvals.
Yogurt is a free, open-source TypeScript app that takes private AI meeting notes, keeping transcripts and summaries local rather than sending them to a hosted service. It drew 133 GitHub stars in its first week.
LASIK is a static CLI and MCP server that reverses Blind LLM Agents hardening so coding agents can read real code again, without executing the target. It reached 100 GitHub stars in its first week.
Huawei-CSL's MACaQ is a model-agnostic quantization method for GGUF and local inference that claims to beat Unsloth GGUFs, letting users quantize and run text, image, video and audio models with llama.cpp, vLLM, Diffusers and PyTorch. llama.cpp adds an FP8 KV-cache.
Brain4Agent is a Python memory module designed to give mainstream AI agents long-term memory tuned for Chinese, Japanese and Korean text; the repo drew 100 stars in its first week.
lot-map is an open-source TypeScript tool that puts a live map beside the Claude Code chat, lighting up tool calls, model steps, hooks, skills and subagents as they happen. Runs locally in five languages.
Pydantic AI is a Python framework for building AI agents with typed models, realtime voice, image generation and embeddings. Versions 2.67.0 and 2.68.0 add SubAgents support under durable execution and strip default-ignorable characters in ArtifactScanner.
Google released Gemini CLI 0.68.0, its open-source terminal agent for Gemini, with changelog additions and CLI fixes including bounds on inline history handling.
An open-source Python project that turns a book into an Agent Skill, using a dual-engine pipeline, pluggable genre plugins and word-level fidelity verification.
Clipto enables AI assistants like ChatGPT and Claude to search and analyze years of a person's meetings and video calls, filling a gap in personal data access for AI assistants.
Time explains how patients can use AI tools to fight back against health insurers that use AI to deny care, offering guidance on appealing denials.
Tether launched BotPsy, an on-device AI agent it claims beats a larger Qwen model, though Qwen's own score narrows the gap.
Pixel 10 owners missing Gboard's Rambler feature may need to force a manual Android AI Core update to enable it.
Tuya Cloud Devices, an open-source skill named gadget-tuya-cloud, lets Meta's Muse AI agent query and control Tuya smart home devices via cloud APIs without local keys or firmware changes.
Manus, a general AI agent that performs actions rather than just generating text, updates its iPhone/iPad app to version 26.19.12 with Manus 2.0 Remote Control for controlling a computer or cloud computer from the phone, plus Manus Studio collaboration.
ByteDance's Doubao AI chat assistant updated to version 2.35.12, available on macOS, Windows and Linux.
Adobe AI Assistant brings conversational AI into Creative Cloud: describe a goal and it handles multi-step creative tasks across Photoshop, Premiere, Illustrator, InDesign, After Effects and Firefly.
Google's Gemini assistant can now operate food-ordering and ride-hailing apps for users in three additional countries.
A report says ByteDance is preparing to release an AI agent that goes beyond a mobile assistant.
Philippine funeral firm St. Peter ran an 'If Heaven Could Write' online campaign from Oct. 27 to Nov. 2, giving customers a digital, interactive, AI-assisted experience of receiving a letter from Heaven.
Tencent's Yuanbao Search, at so.qq.com and in QQ Browser, gives AI answers over the web, WeChat Official Accounts, Channels and academic papers, with cited sources and collapsible ads.
Granola, an AI-powered notepad that transcribes and summarizes meetings, released version 7.786.3 for macOS, Windows and Linux.
Gamma is an AI design tool that turns a prompt into a presentation, document, website or social post. Version 1.1.3 for iPhone and iPad adds new consent options in settings.
Google added Gemini-powered writing tools to Gmail in Workspace, letting users draft and refine emails with AI assistance.
DouBao Work, ByteDance's office AI product, broadened multi-model support as competition among office AI agents intensifies.
In iOS 27, Apple Intelligence features for HomeKit cameras will require higher-storage iCloud+ plans, making AI camera capabilities a paid perk of the subscription service.
Google's NotebookLM-derived Gemini Notebook app for iPhone and iPad reached version 1.66.8, a maintenance release with bug fixes and performance improvements for the research and note-taking assistant.
AI music generator Suno released version 1.97.2 of its iPhone and iPad app, a maintenance update with bug fixes and performance improvements; no new features were announced.
Google Home update restores classic timer sounds, fixes Nest camera Live View issues and improves Gemini commands, faster timers and Nest Hub support.
Sorce, a Y Combinator Fall 2025 startup, lets users swipe right on jobs and sends an AI agent to submit applications. It says 850,000 people have used the tool since its August 2024 launch.
Simpro launched JustDo, an AI agent for trade businesses such as plumbing and electrical contractors, aimed at handling everyday job and business tasks.
Tabletop is a Google Labs experiment where groups build interactive itineraries, polls and brackets with an AI agent. Users can join a waitlist.
On Veterans Day, Echoes Beyond the Wall and Abiogenic.ai launched a humanistic AI effort to honor the 58,281 fallen Vietnam Veterans.
The Llama App, an app for running Meta's Llama models locally, is listed for free download from the Microsoft Store for Windows.
AWS Machine Learning · 23m
#
AWS explains exposing an Amazon Bedrock Managed Knowledge Base to AI agents as MCP tools via AgentCore Gateway, with request interceptors, policy validation and resource-based policies enforcing per-user access.
AWS Machine Learning · 24m
#
AWS explains Agent DLC, a lifecycle approach for building and evaluating production AI agents, covering three agent properties that break traditional CI/CD and a case where an aggregate metric hid a coverage gap. Also covers composable multi-agent systems with A2A on AWS.
KORA Systems, a Silicon Valley startup, is building an AI-centric operating system with Qualcomm, planning a PC OS launch in 2027 before expanding to smartphones.
Databricks Genie, an AI cowork tool that answers business questions over company data with governance controls, is now generally available inside Slack so teams can query data where they work.
At Baidu World 2026 on Nov. 12, Baidu launched Token Exchange, a marketplace for comparing and buying AI model services, alongside model, agent, cloud and chip upgrades; it averages 50M+ calls.
Amp is a frontier agent and development environment; its CLI can be installed via npm with 'npm i -g @ampcode/cli'. This is a new version release.
A guide explains low-rank adaptation (LoRA), which fine-tunes a frozen model by learning small low-rank update matrices instead of changing every parameter, covering trade-offs, evaluation and controls.
Concentrate AI launched a HIPAA-compliant AI gateway that uses a single BAA to simplify model access for healthcare organizations and reduce inference costs.
The latest FreeBSD report shows only a minority want to ban AI contributions, with updates on 16 projects as the community's AI debate continues.
The Linux Foundation added the Agent Name Service, which provides discoverability for agents operating across the internet, backed by 11 companies including Cisco and Salesforce.
A 3D interactive world styled as an N-gauge model railway of Fürth and the line to Munich, built with Claude; every train, bus and tram runs on real timetable and sensor data, with nothing hand-animated.
AI agents are changing how enterprises create, move and govern data, making observability a prerequisite rather than optional for production systems.
An ML engineer released KQuant, a new MLX quantization technique claiming 3.7x better quants for Qwen Flash Next 180B and Qwen 27B, 35B and 9B models, measured by KL-divergence reduction.
OpenAI publishes a customer story on Holiday Extras using Codex to turn customer expertise into internal apps, improve booking conversion and explore new growth areas.
WordPress owner Automattic launched Spacefast, a platform that turns AI-generated projects into working apps and sites, bundling databases, authentication, file uploads, email and analytics.
Harness Memory plugs a chosen memory engine (Mem0, Graphiti, Hindsight, ReMe or custom) into agent harnesses like Claude Code and Codex, sharing memory across harnesses with access control; open-source Community Edition.
Spacefast, by Automattic, lets you connect ChatGPT, Claude, Codex or Cursor and publish a website or app by request, handling hosting, databases, domains, uploads and version history.
A how-to for creating an internal AI support agent in Microsoft Copilot Studio that answers HR, IT and policy questions.
AWS released an open-source framework for evaluating AI agents and hardened its AI grader against prompt-injection attacks.
Cursor's changelog lists fixes and issues: WSL canvas load errors from Windows path backslashes, Cloud Agents dying after first reply since Nov 11, and local MCP running over logged-in ChatGPT, Claude and Grok browser tabs without vendor API keys.
ServiceNow wants Moveworks agent grades to become training rewards, using multi-turn simulations that score each tool call and final system state. It also takes aim at CRM switching costs with its Autonomous CRM Engine.
An essay argues AI agents can deliver working code yet pass along the wrong understanding of why it was built, using a workout-app example.
LightSite AI reports that many websites fail to answer requests from AI agents properly, highlighting a gap as agentic web traffic grows.
Glean posted a job listing for an engineer to use evaluation results and production traces to improve routing, prompts, retrieval and the agent harness, with model post-training handled separately.
The ComfyUI team optimized local LLM inference, making text generation twice as fast in ComfyUI, targeting consumer hardware like the 12GB RTX 3060 for creative workflows.