Perplexity Computer adds Multiagent with up to 16 agents
Perplexity adds a Multiagent effort level above Ultra to Computer: a lead agent runs up to 16 agents on Claude, GPT and Gemini models that share findings and cross-check each other's work.
Perplexity adds a Multiagent effort level above Ultra to Computer: a lead agent runs up to 16 agents on Claude, GPT and Gemini models that share findings and cross-check each other's work.
Bitwise debuted ZDAY, its first non-crypto ETF, betting that spending on AI security will surge.
PopNow AI launched PopNow Protect, which it calls the world's first AI studio with automatic price protection for creative work.
Why it matters: Cache-aware routing can cut inference latency and cost for production LLM endpoints.
Amazon SageMaker AI real-time endpoints now support KV cache-aware routing, sending each request to the instance holding the longest cached match for its prompt, as an alternative to prefix-aware routing.
Why it matters: Adds a multi-model routing and failover layer competing with OpenRouter-style gateways.
Perplexity's Orchestrator API, launched November 17, 2026, gives access to proprietary and open-weight models on its inference infrastructure with consolidated billing, same-model provider failover and preview automatic routing.
Why it matters: Pushes prompt-to-app generation into native Android development with backend and publishing built in.
Google AI Studio now lets developers build, test and publish full-stack Kotlin Android apps from a single prompt, with Cloud Firestore, Firebase Authentication and AI Logic, a faster HD emulator, screenshot debugging and direct Google Play publishing.
Why it matters: Pushes AI video from clips toward consistent multi-episode series production.
Alex Mashrabov's Higgsfield announced a script-to-series tool on Nov. 15 that generates recurring casts, locations and episodes of AI video from a single upload.
Why it matters: Budget-capped agent swarms give autonomous coding a concrete cost guardrail.
Replit's Missions lets users set a goal, success criteria and spending limit, then a swarm of agents plans, builds, tests and iterates until a judge signs off or the budget runs out.
Zenity extended its AI agent security product to cover agents built across the Microsoft ecosystem, adding visibility and controls for enterprise agent deployments.
Why it matters: Auto-approved MCP servers are a real attack surface for coding agents; the fix closes an impersonation path.
Google merged a Gemini CLI fix on November 16 that checks ownership and permissions for system MCP commands, scripts, symlink targets and parent directories, closing an auto-approved MCP server impersonation hole.
SetShot is a Mac utility for guarding against settings that silently change during OS updates. Developer Adam Engst built version 1.0 over months with Claude Code.
Dell launched a gateway that sits between enterprise AI agents and the data they need, aiming to give agents trusted, governed context rather than raw GPU capacity.
Why it matters: The IPO would be a landmark test of public-market appetite for a frontier AI lab.
Reports say Anthropic eyes an IPO price above $80 a share with heavy demand; Amazon and Google would together own nearly 30% afterward, and Ondo offers non-US investors on-chain exposure.
Why it matters: Meta's Muse agent expands its consumer AI assistant footprint into a new market and wearables.
Meta brought its Muse AI agent to Australia with local connectors on mobile, web and Mac, added voice to Apple Watch and Wear OS, and was sued by a small business owner whose own Muse agent helped him.
Why it matters: Microsoft pushes its own specialized models to undercut OpenAI and Anthropic on price.
Microsoft launched new in-house models and agentic continuous-improvement updates, says a new coding model rivals Claude Opus 5.5 at 5% of the cost, and is adding three Meta AI models to Foundry.
Sqwish Labs offers a decision model that companies can train on their own data to improve its capabilities, as adoption of decision models grows.
OpenTSLM, an open-source family of time-series language models, is now in Amazon SageMaker JumpStart, alongside new SageMaker AI data preparation tools for building model-training datasets.
Why it matters: A top hyperscaler publicly commits to data center spending despite political backlash.
Amazon CEO Andy Jassy said at Semafor's Silicon Valley & The World event that pulling back on AI buildout would be a 'historic mistake' and that Amazon expects to hold Anthropic stock long-term.
Lumo is Proton's privacy-focused ChatGPT alternative. The new desktop app for Windows, Mac and Linux automates tasks across files and connected apps, with granular controls over what the AI can access.
Google released ABFS, an open-source content-addressable filesystem and build cache that speeds up Android checkouts to under 8 seconds and full builds to about 2 minutes.
Why it matters: A document-retrieval API aimed at feeding clean web and PDF data to AI agents and RAG pipelines.
You.com released the Extract API on November 17, 2026: one endpoint fetches a web page or PDF and returns Markdown or HTML. Its default PDF mode scored at least 80% higher than Exa Contents and Parallel Extract in a five-dimension test.
Why it matters: More natural script-aware avatars make synthetic presenter video more usable at scale.
HeyGen introduced Avatar 6, an audio-driven, context-aware model that generates 2K avatar video with script-aware gestures and less repetitive, unnatural motion for AI-generated talking-head clips.
Why it matters: A regulatory probe of the top AI chipmaker plus a leadership vacuum adds major uncertainty for AI infrastructure.
The WSJ reports an SEC investigation into Nvidia following Jensen Huang's exit; Nvidia posted 103% revenue growth in its first earnings since his departure and gave a solid outlook.
Datasette Lite runs the Datasette Python app directly in the browser via WebAssembly and Pyodide; a rewrite using service workers makes it nearly identical to the server version.
Willison's post covers OpenAI's GPT-6.2 Astra and Sol launch and notes that LLM 0.38 adds support for the new models in his CLI tool.
Samsara launched a European operations package at Go Beyond Europe 2026 led by Agent Studio, a no-code workspace for building AI agents for driver safety, maintenance, shipment tracking and compliance.
Savana's Acta is operational AI that automates complex banking tasks end-to-end, aiming to take manual back-office work out of bank operations.
A Hackster.io project by Francesco Azzola connects an LLM agent to real environmental sensors using a Raspberry Pi, an ESP32 and the Model Context Protocol, letting agents read sensor data as tools.
Why it matters: Stylist-style agents show how AI is moving from product discovery into completing purchases.
Daydream's AI shopping agent remembers your style, virtually tries on clothes, plans outfits around your wardrobe, tracks sales and completes purchases.
tracee launched applied AI advisory services aimed at regulated financial institutions, helping them adopt AI in line with compliance requirements.
Mural's November 17 release adds an MCP connector, sticky-note conversion and ticket integrations so AI assistants work inside its shared canvas. Mural Compose is in private preview; meeting summaries are in public preview.
Why it matters: Kling extends from video into high-resolution image generation.
Kling AI released Kling Image 4.0, an image generation model it touts for photographic realism and native 8K generation.
Fortra's AI module for GoAnywhere MFT and Globalscape EFT is available to select customers for testing, helping admins with setup and troubleshooting. General availability is expected in early 2027.
TrnRT is an open-source inference engine for running large language models on AWS Trainium chips, giving developers a self-hosted alternative to managed serving on the custom silicon.
Coinbase lets ChatGPT users approve and execute crypto trades through a user-consent flow, keeping confirmation on Coinbase while the assistant handles the request.
An open 9B language model that runs 24 of its 32 stored layers twice, doing ~56 layers of work per pass. Full release scores 36.67 on AutomationBench and 69.90 on GAIA, up from the preview.
Why it matters: Portable memory could break the lock-in that keeps users tied to one assistant.
First Person Labs, founded seven months ago by two PhD students, is building highly personalized AI with portable memory that carries user context across models. Andreessen Horowitz backs the startup.
Why it matters: Shows frontier models can be nudged toward lethal targeting errors, a serious military-AI safety concern.
An independent research report found advanced AI models in simulated scenarios regularly misidentified civilians, ambulances, surrendering soldiers and a school in Iran as military targets, and can be tricked by framing it as a game.
The startup from a Google Maps alum raised $12M to build an AI agent that books travel, betting a specialized travel planner beats general-purpose assistants.
Sarvam AI launched a Token Factory service in India that routes and manages model inference tokens, positioning itself against OpenRouter for local developers.
Dell says rising AI factory density makes storage the bottleneck, and is turning PowerEdge servers into liquid-cooled, software-defined storage nodes that can switch roles as workloads shift.
Tesla says the US-made unit combines transformer, inverter and breaker for storage, solar and 800-volt data centers, with first operation targeted for 2027.
Why it matters: Eases programming of wafer-scale hardware and pairs it with GPU disaggregation.
Cerebras released Tungsten and Aviary, tools that remove assembly-like work needed to program its 900,000-core wafer-scale systems, and detailed handing prefill to AMD GPUs for more tokens per megawatt.
Why it matters: A regulator grounding freeway robotaxis is a major test of self-driving safety and expansion.
California's DMV ordered Waymo off state freeways, calling the robotaxis an imminent hazard after construction-zone incidents; Waymo paused all freeway rides. Hyundai will start delivering Waymos in December.
A UK defense company developed a low-cost radar that detects drones and aircraft from an SUV without revealing its own location.
Why it matters: 800 VDC distribution at pod and hall scale is a key enabler for denser, higher-power AI datacenters.
Vertiv introduced PowerDirect 9000 and PowerNexus MV SST, pushing 800-volt DC power beyond single racks to AI compute pods and data halls; PowerNexus is a 5 MW solid-state transformer converting medium-voltage AC.
A TensorRT-LLM change adds an opt-in MXFP8 kernel targeting replicated shared experts on SM107; it is off by default and reached 2.5x in a Kimi K3 microbenchmark on one VR200.
Databricks says a new version of Azure Databricks built for Azure Government is in beta, aimed at agencies running data and AI workloads on classified mission systems.
Edifier released the Huazai Halo PixelBar AI, a desktop soundbar that sits under a monitor, combining a pixel display and dual 52mm drivers. It costs 699 yuan ($104) and is on sale on JD.com.
Uber Eats has launched drone delivery in a Texas suburb in partnership with Zipline, bringing autonomous aerial delivery to everyday food orders.
EssilorLuxottica and Meta launched their Ray-Ban Meta AI glasses in South Africa, backed by a holiday 'Window Shopping' ad campaign framing the glasses as gifts.
Why it matters: Nvidia is positioning agentic AI and a 2028 HPC chip at the core of scientific computing.
At SC26 Nvidia announced Rubin HPC, a science-focused server GPU due in 2028, Nemotron for Science agent tools, and software optimizations that lift its Vera CPU up to 39%.
Why it matters: ASUS joins the humanoid race with a hospital-focused robot, extending consumer tech firms into healthcare robotics.
ASUS introduced Gallop, its first humanoid robot, designed for hospital environments.
US firm Aviemore Motion introduces a seal-less magnetic pump aimed at leak-free liquid cooling for AI data centers.
Why it matters: Adds a photonic computing option to the LLM accelerator race beyond GPUs.
German photonics company Q.ANT unveiled its NPU Gen 2+, an optical neural processing unit that uses photonic Fourier transforms to accelerate large language model workloads.
Rockwell Automation expands its edge computing options for industrial AI by adding Nvidia GPUs.
Verso is a MagSafe AI voice recorder with an E-Ink display that attaches to a phone; Flowtica has revealed the device.
Why it matters: Gives GPU operators an open way to run managed Kubernetes on bare metal.
Nvidia open-sourced NVKE and NVKR under Apache 2.0, letting AI factories run managed Kubernetes on their own bare-metal GPU servers; both are in public preview. It is also rewiring chip interconnects with light.
At Supercomputing 2026, Silicon Motion showcased next-generation PCIe Gen5 and Gen6 enterprise storage for AI factories and agentic AI, plus an enterprise boot drive paired with an Nvidia DPU.
Why it matters: AMD is mounting a direct challenge to Nvidia in AI inference and HPC accelerators.
AMD says its Instinct MI455X in the Helios system hit 21,333 tokens per second, 30% above Nvidia, touts native FP64 on MI430X for HPC, and closed a Taalas deal for model-specific inference chips.
ZutaCore launched Z2O, a cooling system it says eliminates water use in AI data centers, targeting the power and water constraints of dense GPU racks.
IEI Integration and medical unit BriteMED are showing edge AI computing and smart healthcare hardware, including medical workstations, at MEDICA 2026 in Düsseldorf.
Abu Dhabi investment firm Lunate is launching the Boreas VettaFi Humanoid Robots UCITS ETF, listing on the Abu Dhabi Securities Exchange on December 8, giving GCC investors humanoid robot exposure.
Noetix Robotics released an over-the-air V3.2.0 software update for its Bumi robot.
Why it matters: Local opposition is emerging as a real constraint on AI data center buildouts.
Giga Energy signed a $1 billion lease for a Texas AI data center with a cloud provider, as voter backlash threatens Amazon's $17 billion Ohio data center bet.
Factory's AI software engineering agent ships version 1.13.1 across macOS, Windows and Linux.
Perplexity adds a Multiagent effort level above Ultra to Computer: a lead agent runs up to 16 agents on Claude, GPT and Gemini models that share findings and cross-check each other's work.
AWS describes three composable strategies for Amazon Bedrock AgentCore agents handling multi-megabyte JSON: offloading artifacts to S3, on-demand result paging and dictionary compaction, cutting per-turn token use.
Antigravity, an agent orchestration platform, shipped version 2.32.2 for macOS, Windows and Linux.
AnythingLLM, a private desktop app for chatting with your own documents and models locally, released version 1.20.0 for macOS, Windows and Linux.
Codex CLI, OpenAI's local coding agent for the terminal, released version 0.169.1, installable via npm.
A Fello AI post covers MAI Code 2, detailing its pricing, benchmark results and planned integration into Microsoft Copilot.
Docusign's ChatGPT plugin now supports sales renewal preparation, legal-request routing and agreement generation. The plugin is available for download and also works with OpenAI's Sales plugin.
Fathom is an app that records and transcribes video conferences. Version 3.14.1 is a maintenance update for macOS, Windows and Linux.
Why it matters: Cache-aware routing can cut inference latency and cost for production LLM endpoints.
Amazon SageMaker AI real-time endpoints now support KV cache-aware routing, sending each request to the instance holding the longest cached match for its prompt, as an alternative to prefix-aware routing.
PCMag looks at a $49 tool that shows how ChatGPT, Claude and Gemini describe a given brand in their answers.
MarkTechPost guide explains that zero data retention applies to specific AI endpoints rather than whole vendors, compares two standard API retention designs, who handles deletion, and six checks for developers.
Why it matters: Adds a multi-model routing and failover layer competing with OpenRouter-style gateways.
Perplexity's Orchestrator API, launched November 17, 2026, gives access to proprietary and open-weight models on its inference infrastructure with consolidated billing, same-model provider failover and preview automatic routing.
SiliconANGLE explains how agentic workloads that sit idle then spike unpredictably strain static Kubernetes resource settings and push platform teams toward autonomous orchestration.
Administrators can define validated metadata fields every agent, MCP server and skill must carry, then follow a record from schema definition through registration, approval and discovery.
Kaseya added MCP servers and AI skills to its agentic IT management platform, extending what its agents can connect to and automate for IT teams.
iOS 27.3 will bring an all-new Apple Health experience next month, but the redesign requires a newer, AI-compatible iPhone model to run its features.
Why it matters: Pushes prompt-to-app generation into native Android development with backend and publishing built in.
Google AI Studio now lets developers build, test and publish full-stack Kotlin Android apps from a single prompt, with Cloud Firestore, Firebase Authentication and AI Logic, a faster HD emulator, screenshot debugging and direct Google Play publishing.
Why it matters: Budget-capped agent swarms give autonomous coding a concrete cost guardrail.
Replit's Missions lets users set a goal, success criteria and spending limit, then a swarm of agents plans, builds, tests and iterates until a judge signs off or the budget runs out.
OpenAI's November 16 SDK update reverses the documented background default to blocking compaction, adds target_session_failed for paused schedules, and adds a Live safety identifier.
A merged vLLM change adds A100 support to the decode path for DeepSeek V4.1-Flash, with the contributor reporting 488.8 P90 tokens per second per user on H100 under different configurations.
SetShot is a Mac utility for guarding against settings that silently change during OS updates. Developer Adam Engst built version 1.0 over months with Claude Code.
Dell launched a gateway that sits between enterprise AI agents and the data they need, aiming to give agents trusted, governed context rather than raw GPU capacity.
Why it matters: Meta's Muse agent expands its consumer AI assistant footprint into a new market and wearables.
Meta brought its Muse AI agent to Australia with local connectors on mobile, web and Mac, added voice to Apple Watch and Wear OS, and was sued by a small business owner whose own Muse agent helped him.
Tests show Gemini for Home mishandles basic Spotify song, artist and playlist requests despite an August update meant to improve recognition. Google confirmed a large list of new Home features and three Nest Thermostat upgrades.
OpenTSLM, an open-source family of time-series language models, is now in Amazon SageMaker JumpStart, alongside new SageMaker AI data preparation tools for building model-training datasets.
The Register looks at the open-source Crosspoint Reader e-reader project and how it handles drive-by pull requests that appear generated by large language models.
Lumo is Proton's privacy-focused ChatGPT alternative. The new desktop app for Windows, Mac and Linux automates tasks across files and connected apps, with granular controls over what the AI can access.
Manus AI is a GitHub project offering a general-purpose AI agent system that runs tools and operations inside a sandbox. The repo drew 202 stars in its first week.
Google released ABFS, an open-source content-addressable filesystem and build cache that speeds up Android checkouts to under 8 seconds and full builds to about 2 minutes.
Multimodal Knowledge Agent (MKA) is a Google DeepMind benchmark dataset for evaluating autonomous AI agents on complex, multi-step visual knowledge work. Each of its 10 showcase tasks uses a multi-file workspace that plain OCR or text extraction cannot solve.
USDCraft-10K is a dataset of simulation-ready rigid and articulated 3D assets in OpenUSD, built by Nvidia with a method where a language model writes each object as an asset program that compiles into parts, joints and kinematics.
A LoRA adapter on Qwen3.8-27B that makes the model more efficient when run inside the Pi agent harness; it can be run locally, part of Hugging Face's harness-adapters work.
tiny-voca is a 4.5M-parameter English text-to-speech model with two voices that runs about 70x faster than real time on four CPU threads; it ships alongside KLM-1.6, Nex-N2.6 and Edge0-145B-q3k releases on Hugging Face.
A dataset of Blender project files and videos behind 154 artworks by ARC, made 2019-2024, released under CC BY-NC 4.0 for non-commercial learning and remixing.
K2-Horizon is a family of six fully open language models, from a 0.9B dense model to a 375B-A23B MoE, built for reasoning and agentic work across coding, math and science.
An open medical AI evaluation initiative whose Atlas maps benchmarks across three healthcare roles and three AI capabilities, with four benchmarks in the first release.
A paper asks whether frontier-scale vision-language representations improve robot control and scales vision-language-action policies to 552 billion parameters, far larger than typical VLA backbones.
Manus AI is a GitHub project offering a general-purpose AI agent system that runs tools and operations inside a sandbox. The repo drew 202 stars in its first week.
An agent skill that hill-climbs local LLM serving speed across decode, speculative decoding, prefill, KV cache, startup and TTFT on any hardware and quantization. The repo drew 478 stars in its first week.
An MIT-licensed Python repo collecting 31 marketing skills an AI agent can use for research, positioning, copy, SEO, AI search, ads and email, plus a local-business pack.
Goose is an open-source, extensible AI agent from Block that installs, executes, edits and tests code with any LLM. Version 1.60.0 opens a new release branch.
A GitHub project that records verified license and provenance data for medical AI datasets, using information cards, a JSON Schema, decision rules and a hash-only evidence index. Cards are CC BY 4.0, scripts MIT.
Gemini 4 Hydrogen is a fast, efficient Google model built for robotics, with spatial and embodied reasoning. A new Robotics plugin for Google Antigravity turns it into a workbench for building and testing robotic agents.
turbo-browser-mcp is an open-source Playwright MCP server that lets AI agents control a browser with fewer turns and tokens, via one-call form filling, delta snapshots, batched steps and an on-by-default submit guard.
vLLM is an open-source high-throughput inference and serving engine for LLMs. Version 0.34.0 cuts a DeepSeek V4.1 Engram lookup from 54 ms to 5.1 ms on eight MI355X GPUs.
Perplexity adds a Multiagent effort level above Ultra to Computer: a lead agent runs up to 16 agents on Claude, GPT and Gemini models that share findings and cross-check each other's work.
AnythingLLM, a private desktop app for chatting with your own documents and models locally, released version 1.20.0 for macOS, Windows and Linux.
Docusign's ChatGPT plugin now supports sales renewal preparation, legal-request routing and agreement generation. The plugin is available for download and also works with OpenAI's Sales plugin.
Fathom is an app that records and transcribes video conferences. Version 3.14.1 is a maintenance update for macOS, Windows and Linux.
PCMag looks at a $49 tool that shows how ChatGPT, Claude and Gemini describe a given brand in their answers.
iOS 27.3 will bring an all-new Apple Health experience next month, but the redesign requires a newer, AI-compatible iPhone model to run its features.
SetShot is a Mac utility for guarding against settings that silently change during OS updates. Developer Adam Engst built version 1.0 over months with Claude Code.
Meta brought its Muse AI agent to Australia with local connectors on mobile, web and Mac, added voice to Apple Watch and Wear OS, and was sued by a small business owner whose own Muse agent helped him.
Tests show Gemini for Home mishandles basic Spotify song, artist and playlist requests despite an August update meant to improve recognition. Google confirmed a large list of new Home features and three Nest Thermostat upgrades.
Lumo is Proton's privacy-focused ChatGPT alternative. The new desktop app for Windows, Mac and Linux automates tasks across files and connected apps, with granular controls over what the AI can access.
Ambience Healthcare added Rounding Assistant and Discharge Readiness to its inpatient AI suite, prioritizing tasks and flagging discharge barriers earlier in hospital workflows.
Anthropic shipped Claude desktop app version 6.6395.0 for macOS, Windows and Linux and launched a beta of linked billing for Claude API apps.
Speechify is an AI reading and voice assistant that turns documents, web pages and PDFs into spoken audio. Version 3.25.0 is out for macOS, Windows and Linux.
The startup from a Google Maps alum raised $12M to build an AI agent that books travel, betting a specialized travel planner beats general-purpose assistants.
AI Playground is a free Windows app from Microsoft Research for trying out local AI models on your own PC. It lets you run chat, image generation and other models locally instead of in the cloud.
Gemini can now send texts and emails and call businesses to complete purchases via Channels, while Google Maps launched AI itineraries and expanded Ask Maps across Europe.
Doubao, ByteDance's AI chat assistant, released version 2.36.10 for macOS, Windows and Linux.
Samsung is testing 'Agentic Home' features in SmartThings to run smarter automations and cut energy bills, and references to paid Galaxy AI tiers suggest extra Now Brief custom cards.
KAYAK rolled out 28 new and enhanced AI features across its travel search, expanding Ask AI with flexible-date price searches, conversational itineraries, mobile access and deeper flight and hotel results.
Google's Spark AI agent is now 'Tasks' and free, with wider access for calls and purchases; Gems retire today and Opal shuts down Nov. 17 without automatic app transfer.
MacWhisper is a speech recognition tool that transcribes audio on your machine. Version 15.11 is out for macOS, Windows and Linux.
Apple released a research study and iPhone app probing how people use AI health tools. The app has a generative AI chatbot that answers health questions with peer-reviewed citations and pulls the user's own health data.
Typeless is an AI voice dictation app that turns spoken words into polished written text. Version 2.12.0 ships for macOS, Windows and Linux.
A 16-year-old in Karachi built Smartpet, a voice-guided AI learning assistant that breaks any goal into ten steps for beginners. He applied to Y Combinator with it.
Link.bio generates a personalized link-in-bio page from a pasted Instagram, TikTok, YouTube or site link: AI writes the bio, buttons and theme, and the page is then run by an agent.
Factory's AI software engineering agent ships version 1.13.1 across macOS, Windows and Linux.
AWS describes three composable strategies for Amazon Bedrock AgentCore agents handling multi-megabyte JSON: offloading artifacts to S3, on-demand result paging and dictionary compaction, cutting per-turn token use.
Antigravity, an agent orchestration platform, shipped version 2.32.2 for macOS, Windows and Linux.
Codex CLI, OpenAI's local coding agent for the terminal, released version 0.169.1, installable via npm.
A Fello AI post covers MAI Code 2, detailing its pricing, benchmark results and planned integration into Microsoft Copilot.
Amazon SageMaker AI real-time endpoints now support KV cache-aware routing, sending each request to the instance holding the longest cached match for its prompt, as an alternative to prefix-aware routing.
MarkTechPost guide explains that zero data retention applies to specific AI endpoints rather than whole vendors, compares two standard API retention designs, who handles deletion, and six checks for developers.
Perplexity's Orchestrator API, launched November 17, 2026, gives access to proprietary and open-weight models on its inference infrastructure with consolidated billing, same-model provider failover and preview automatic routing.
SiliconANGLE explains how agentic workloads that sit idle then spike unpredictably strain static Kubernetes resource settings and push platform teams toward autonomous orchestration.
Administrators can define validated metadata fields every agent, MCP server and skill must carry, then follow a record from schema definition through registration, approval and discovery.
Kaseya added MCP servers and AI skills to its agentic IT management platform, extending what its agents can connect to and automate for IT teams.
Google AI Studio now lets developers build, test and publish full-stack Kotlin Android apps from a single prompt, with Cloud Firestore, Firebase Authentication and AI Logic, a faster HD emulator, screenshot debugging and direct Google Play publishing.
Replit's Missions lets users set a goal, success criteria and spending limit, then a swarm of agents plans, builds, tests and iterates until a judge signs off or the budget runs out.
OpenAI's November 16 SDK update reverses the documented background default to blocking compaction, adds target_session_failed for paused schedules, and adds a Live safety identifier.
A merged vLLM change adds A100 support to the decode path for DeepSeek V4.1-Flash, with the contributor reporting 488.8 P90 tokens per second per user on H100 under different configurations.
Dell launched a gateway that sits between enterprise AI agents and the data they need, aiming to give agents trusted, governed context rather than raw GPU capacity.
OpenTSLM, an open-source family of time-series language models, is now in Amazon SageMaker JumpStart, alongside new SageMaker AI data preparation tools for building model-training datasets.
The Register looks at the open-source Crosspoint Reader e-reader project and how it handles drive-by pull requests that appear generated by large language models.
Google released ABFS, an open-source content-addressable filesystem and build cache that speeds up Android checkouts to under 8 seconds and full builds to about 2 minutes.
You.com released the Extract API on November 17, 2026: one endpoint fetches a web page or PDF and returns Markdown or HTML. Its default PDF mode scored at least 80% higher than Exa Contents and Parallel Extract in a five-dimension test.
Datasette Lite runs the Datasette Python app directly in the browser via WebAssembly and Pyodide; a rewrite using service workers makes it nearly identical to the server version.
Willison's post covers OpenAI's GPT-6.2 Astra and Sol launch and notes that LLM 0.38 adds support for the new models in his CLI tool.
A Hackster.io project by Francesco Azzola connects an LLM agent to real environmental sensors using a Raspberry Pi, an ESP32 and the Model Context Protocol, letting agents read sensor data as tools.
Mural's November 17 release adds an MCP connector, sticky-note conversion and ticket integrations so AI assistants work inside its shared canvas. Mural Compose is in private preview; meeting summaries are in public preview.
TrnRT is an open-source inference engine for running large language models on AWS Trainium chips, giving developers a self-hosted alternative to managed serving on the custom silicon.