Why it matters: A cheaper, faster coding model lowers cost for routine autonomous software-agent tasks.
Cognition's SWE-2 Flash coding model runs three times faster and at one-fifth the price of SWE-2, available in Devin Desktop and CLI for routine software-agent work.
BeyondTrust launched AI agent security capabilities, released with accompanying research, aimed at securing enterprise AI agents.
Forter introduced new capabilities built on an agentic commerce engine, per Retail TouchPoints, extending its fraud and trust platform to AI shopping agents.
MariaDB released Enterprise Platform 2027, generally available, combining transactional, analytics, in-memory and AI capabilities with agent memory, retrieval and response caching; peak throughput up 56%.
BotGauge launched an AI agent red-teaming platform with over 200 attack strategies, plus tracing and evaluation tools for testing agent security.
Anthropic added partner agents and products to the Claude Marketplace, and tweaked how Claude Code handles requests about making games.
Why it matters: A 65B open model claiming US open-weight SOTA on agentic coding signals rising competition among open labs.
Ripple-65B Preview is a 65B dense open-weight model trained on 12 trillion tokens with up to 1M context. Ricursive says it sets a US open-weight state of the art on agentic coding, beating larger models.
Why it matters: Targets cost and governance of enterprise coding agents with model routing.
Baseten launched Grid, a public preview of a routing, gateway and guardrails layer for enterprise AI coding. Baseten says its Router cut cost per task nearly 70% with a 0.9-point quality tradeoff in initial benchmarks.
Xiora AI is a web platform from Xiora Technologies for personalized AI video companions, letting users interact with AI-generated video characters.
Why it matters: A cheap fast tier of a leading AI video model lowers cost for high-volume generation.
Kuaishou's Kling 4.0 Flash video generation model is generally available at $0.094 per second with a 720p resolution ceiling.
Jobber lets eligible US home and commercial service businesses be discovered and booked through Google Search's AI Mode, with requests flowing into existing Jobber workflows and no lead fees.
Contextbase stores AI agent context — semantics, knowledge, memory, traces and evaluations — inside PhoenixAI's database. It is in private preview; ObserveStack is the first app built on it.
Suno now accepts MIDI file uploads in its Create tool, letting users feed melody and note data into AI song generation.
Phone Harness gives AI agents control of a real iPhone or a rented cloud phone. Founder Shawn Pana open-sourced the harness for personal devices and sells persistent cloud phones.
Why it matters: Makes multi-agent coding in a single editor session mainstream for developers.
The new VS Code release lets developers run Claude, Codex and GitHub Copilot agents within the same agent session, per Windows Report.
The Dharma Field Extractor is a 2B model that takes any document and outputs structured fields as JSON. Dharma-AI's Gabriel Pimenta released it on Hugging Face.
Unsloth released Dynamic v3.1 quantized GGUFs for Qwen3.8-Flash-Next, claiming about 15% better accuracy than GSQ RCO at similar file size; users report IQ_3S GSQ RCO nearing BF16 quality.
AT&T's Connected Life home security system now integrates with Gemini for Home, adding Gemini-powered event summaries and video history search. AT&T calls it the first home security product to use Google Home Premium.
Viggle Remake is an open-weights model for recasting people and editing scenes in existing videos while keeping the original motion; weights are on Hugging Face and a demo app is available.
A YC S23 startup launching an AI media manager and video editor that organizes footage and helps edit video; details of features and availability are limited in the post.
Why it matters: Shows a 2B multimodal model running fully offline on a phone at usable speeds.
LFM3-Omni-2B is a small multimodal model that can run entirely on-device. A Hugging Face Space demo and a React Native app using custom WebGPU kernels run it on an iPhone at about 80 tokens per second, fully offline.
Netflix is rolling out a Discover feature that recommends tailored streaming selections to help users find something to watch. ZDNET explains how it works and how to try it.
At ENGAGE 2026, Smartsheet launched vsActual, a standalone AI-native planning and execution product, plus an 'execution graph' aimed at linking team tasks to organizational strategy.
Lorikeet launched Dreamweaver, an AI account manager built into its Concierge agents that pursues activation and retention goals across voice, SMS and email; early results cited at PayJoy.
OpenEvidence launched Learn, a free personalized medical education platform for trainees and clinicians, extending its AI from point-of-care answers into education, developed with medical institutions.
SkyFi's Satellite Intelligence app is available to Robinhood Agents customers, offering BlackSky observations and AI analysis of 24 sites tied to 10 public companies for $10 a month.
Why it matters: AI video detection is a key defence as generated video becomes harder to distinguish from real footage.
Pangram released two detection models for AI-generated video and images. The company claims its video detector scores 98.7%, and it also updated its existing image detector.
Why it matters: Suggests frontier models can directly reprogram other models, a notable capability for AI research.
MIT researchers found OpenAI's GPT-6 Astra can read and deliberately rewrite other models' weights without traditional training, boosting a Qwen model's math accuracy from 7.9% to 48.1% by changing 0.016% of its weights.
Why it matters: A 2.8T-parameter open-weight release pushes the frontier of freely available models.
Moonshot AI published open weights for Kimi K3.1 Pro, a 2.8-trillion-parameter model, with day-0 support and optimization on AMD Instinct GPUs.
CrowdStrike released an 'AI Kill Switch' product that lets enterprises shut down or halt AI systems and agents, launched as investor attention around Anthropic's reported IPO valuation grows.
Leidos introduced HeadWay Dispatch, software for overseeing and controlling AI agents across different vendors, aimed at organizations running multi-vendor agent fleets.
A personal trainer in Wales built an AI-powered app aimed at improving gym training for users. Further product details were not specified in the report.
About Amazon · 7h · 3 sources
#
Why it matters: Expands how merchants surface products to AI shopping agents, a key agentic commerce battleground.
Amazon's new Shop Direct Merchant Workspace gives merchants a free direct AI connector so their products can be discovered by customers via Alexa for Shopping and Amazon.
Why it matters: Lets websites distinguish and govern autonomous agent traffic instead of treating it as ordinary bot or human visits.
Cloudflare released analytics and per-agent access controls so site owners can see personal AI agents visiting their sites and set terms for what those agents may do.
Adobe integrated its Agentic CX Enterprise Coworker into AI platforms from Anthropic, Google, Microsoft, OpenAI and others.
MyHeritage introduced GAIA, an AI agent that helps users research family history and ancestry records.
ModelOp launched what it calls a full-stack AI policy enforcement offering built with Nvidia's OpenShell, aimed at governing AI model behavior in production.
Atlas is a free, MIT-licensed tool that creates a hidden second desktop on a Mac where computer-use AI agents can click and type while the user keeps working undisturbed.
Cycurion launched Habit, an AI-driven security operations center that learns what normal looks like and hunts anomalies, aimed at automating threat detection and response.
Snowflake launches autonomous AI agents paired with tighter enterprise data controls, letting organizations run agents on their data while keeping governance in place.
Fermaca, designated Mexico's first Nvidia Cloud Partner, announced a phased AI factory platform for Mexico and Latin America, starting with Fermaca Digital Origins in Mexico City.
Why it matters: CUDA support signals Vera Rubin hardware is production-ready and gives developers the toolchain for it.
CUDA 13.5 adds support for Nvidia's Vera Rubin platform reaching general availability, giving developers the toolchain for the next-gen accelerated computing hardware.
An open-source project builds an AI-powered claw suspended from cables near the ceiling, letting it pick things up around a room instead of using a humanoid robot.
Broadcom added AI, security and recovery capabilities to VMware Cloud Foundation, its private cloud platform.
Why it matters: World models are a key path toward training physical AI and robots in simulation.
Baidu introduced a new world model aimed at physical AI applications, with coverage highlighting both its potential and the challenges of applying it to real-world robotics and simulation.
Weebit Nano launched a research program giving academic and industry researchers ReRAM hardware kits built around a 32x32 crossbar to test running AI calculations where data is stored, rather than moving it.
Nvidia released its new R620 Linux graphics driver series, version 620.70.06, with Wayland improvements and DLSS 5 support for games including Hogwarts Legacy and World of Warcraft.
Why it matters: Voter-driven restrictions on data centers could slow the power and construction underpinning AI expansion.
After the Democratic sweep, voters backed candidates favoring moratoriums and ballot measures that slow data center construction. Analysts and outlets including Bloomberg, Wired and the FT see new restrictions on a pillar of the AI boom.
Hyundai Motor Group demonstrates a robot that automatically refuels a hydrogen-powered cleaning truck, alongside a 5 MW hydrogen system designed to produce over 700 tons annually.
Why it matters: Shows Google and Anthropic locking in massive AI compute capacity.
TeraWulf secured a Google credit backstop for its $19 billion Anthropic data-center lease project.
UBTECH and Fenglong Shares' joint venture launched a portfolio of robot joint products, components for humanoid and embodied AI robots.
1Finity released a full-stack virtualized RAN suite that uses Nvidia-powered GPU infrastructure sharing, targeting telecom operators.
Open-source developers released a prototype driver that renders the Windows 11 desktop using the NPU instead of the iGPU or GPU. It currently runs at only 2 to 3 Hz, with a 30 Hz target.
Why it matters: A sub-$8,000 humanoid undercuts typical humanoid pricing and widens access.
StarBot launched Spark, a US-made humanoid robot priced at $7,999, aimed at buyers wanting a humanoid platform at consumer-grade cost.
Nordic Semiconductor announced a new Wi-Fi and edge AI chip aimed at battery-powered IoT devices.
Aetina released a rugged edge AI system built on Nvidia Jetson for AI inference in harsh environments.
Tokyo police plan an AI system combining 3D spatial data from crash sites with security camera footage to reconstruct accidents, reducing the need for on-site inspections.
HONPINE's TCHL joint module integrates a built-in torque sensor aimed at humanoid robots, packaging sensing into the actuator.
Innodisk is supplying enterprise SSDs and AI server memory modules for AI data centers, with its data center SSD series in full customer sampling and DDR5 modules expanding.
PrimeBOT is opening its Q1 Explorer home robot's hardware and software to owners, letting them customize the customizable home robot.
Why it matters: Unifies edge-AI silicon for automotive, industrial and robotics designers.
STMicroelectronics introduced a unified STM32 microcontroller platform and three STM32A product lines for high-safety and automotive systems, extending a shared edge-AI foundation across industrial, automotive and robotics designs.
Why it matters: A 2nm AI chip marks a process-node milestone for AI silicon.
Broadcom has shipped a landmark AI chip built on a 2nm process, but investors gave the news a muted reaction, Startup Fortune reports.
Why it matters: Shows mixed-vendor GPU pooling to reduce dependence on a single accelerator ecosystem.
GPUWorks is an Apache-2.0 inference platform that pools Ascend, T-Head PPU, Hygon and NVIDIA GPUs in a single cluster; developers say one server ran 1,003 GPU machines.
Nvidia App update 11.0.9.585 introduces GPU app prioritisation for some GeForce RTX owners, letting users split graphics resources between programs, plus 19 new DLSS override profiles.
UK insurer YAS and robotics firm Robolist released an insurance policy covering humanoid robots, addressing a gap as more robots enter workplaces and homes without coverage.
Raycast, a keyboard-driven launcher for controlling tools and running commands, released version 2.11.1.0 on macOS, Windows and Linux.
Apple's updated support document says the Siri AI waitlist used to gate access at launch in iOS 27 is no longer in use, except during high-demand periods.
An open-source AI agent that brings Gemini directly into the terminal — v0.67.0 adds an IDE bridge and other changes. Free and open source for developers.
Why it matters: A cheaper, faster coding model lowers cost for routine autonomous software-agent tasks.
Cognition's SWE-2 Flash coding model runs three times faster and at one-fifth the price of SWE-2, available in Devin Desktop and CLI for routine software-agent work.
MariaDB released Enterprise Platform 2027, generally available, combining transactional, analytics, in-memory and AI capabilities with agent memory, retrieval and response caching; peak throughput up 56%.
Oracle is pitching a shared semantic layer and new security and AI-driven testing tools as enterprises pull data back into its database for real-time agent access.
Zed is a fast collaborative code editor; version 1.27.0 ships for macOS, Windows and Linux with its multiplayer editing features.
AWS explains connecting decision models Jev and Strands Decider to Amazon Bedrock AgentCore Evaluations via code-based evaluators, to answer narrow, repeatable agent-evaluation questions.
Why it matters: Targets cost and governance of enterprise coding agents with model routing.
Baseten launched Grid, a public preview of a routing, gateway and guardrails layer for enterprise AI coding. Baseten says its Router cut cost per task nearly 70% with a 0.9-point quality tradeoff in initial benchmarks.
After major Republican election losses, users asked Grok whether Elon Musk deleted his 'Vote Red Today!' post; the xAI chatbot said yes, per Yahoo. An example of AI answering questions about its owner's posts.
ChatGPT can now generate animated GIFs from text prompts or uploaded photos using OpenAI's Images 2.6 model. Users can adjust animation speed, edit frames and share results.
Contextbase stores AI agent context — semantics, knowledge, memory, traces and evaluations — inside PhoenixAI's database. It is in private preview; ObserveStack is the first app built on it.
Phone Harness gives AI agents control of a real iPhone or a rented cloud phone. Founder Shawn Pana open-sourced the harness for personal devices and sells persistent cloud phones.
Why it matters: Makes multi-agent coding in a single editor session mainstream for developers.
The new VS Code release lets developers run Claude, Codex and GitHub Copilot agents within the same agent session, per Windows Report.
AT&T's Connected Life home security system now integrates with Gemini for Home, adding Gemini-powered event summaries and video history search. AT&T calls it the first home security product to use Google Home Premium.
Replay's QA tool surfaced more than 900 bugs across 40 promoted apps, according to CEO Brian Hackett, and traced them to root causes AI agents can act on.
Manus for Mac is an AI agent app that automates workflows on your own computer. Version 2.1.1 is a maintenance update for macOS, Windows and Linux.
Redis's Iris real-time context engine now offers six composable capabilities for AI agents — four generally available, two in beta — adding request routing, reusable procedures, observability and resilient messaging.
ZDNET reports new Safari developer beta release notes point to deeper Siri AI integration and more Safari customization, plus three popular browser features.
Why it matters: Shows a 2B multimodal model running fully offline on a phone at usable speeds.
LFM3-Omni-2B is a small multimodal model that can run entirely on-device. A Hugging Face Space demo and a React Native app using custom WebGPU kernels run it on an iPhone at about 80 tokens per second, fully offline.
Netflix is rolling out a Discover feature that recommends tailored streaming selections to help users find something to watch. ZDNET explains how it works and how to try it.
BoltAI is an AI chat client that lets you use multiple models in one desktop app. Version 2.18.5 is a maintenance release for macOS, Windows and Linux.
At ENGAGE 2026, Smartsheet launched vsActual, a standalone AI-native planning and execution product, plus an 'execution graph' aimed at linking team tasks to organizational strategy.
A framework for production AI agents covering architecture, prompts, file-based tools and verification agents, implemented in working code from Deep Insight, an open-source multi-agent data analysis solution.
Siri Recaps, Apple's AI feature that summarizes audio and notifications, was tested in watchOS 27.3. Reviewers called it useful but raised privacy concerns and noted room for improvement.
Nvidia's NeMo tooling turns production evidence into prioritized issues and custom evaluations to improve prompts, tools, models, memory and harness code for AI agents.
Why it matters: A 2.8T-parameter open-weight release pushes the frontier of freely available models.
Moonshot AI published open weights for Kimi K3.1 Pro, a 2.8-trillion-parameter model, with day-0 support and optimization on AMD Instinct GPUs.
Cohere's Embed 5 and Rerank 5 retrieval models are now on OpenRouter, callable with the same API key and credits as other models without separate SDKs, accounts or bills.
datasette-llm is a plugin that adds LLM features to Datasette; version 0.1a10 fixes a bug where available models were not limited to those in the configuration.
Leidos introduced HeadWay Dispatch, software for overseeing and controlling AI agents across different vendors, aimed at organizations running multi-vendor agent fleets.
LFM3-Omni-2B is a small multimodal model that can run entirely on-device. A Hugging Face Space demo and a React Native app using custom WebGPU kernels run it on an iPhone at about 80 tokens per second, fully offline.
Moonshot AI published open weights for Kimi K3.1 Pro, a 2.8-trillion-parameter model, with day-0 support and optimization on AMD Instinct GPUs.
A model that turns a single image into a 3D asset, a fine-tune of TripoSplat, repackaged with ComfyUI-ready files under Apache-2.0. Drop-in checkpoints plug into existing TripoSplat workflows.
PentesterFlow/OffsecCyber is an Apache-2.0 English dataset of security assistant conversations covering CVE analysis, CWE references, remediation and web-appsec reasoning.
Hydrogen-1.0 is a scientific problem-solving model from Deep Principle built on GLM-5.3 that helps answer scientific questions, explain derivations and organize responses to open-ended research problems.
Idol is a Gradio demo that converts a single image into a 3D object as Gaussian splats, using DINOv3 and FLUX.2-dev image encoders; running it requires an HF token with access to those models.
A community MiniMax-H3 image model 7x smaller and 5x faster in 8 steps, plus PrunaAI's ComfyUI model and a drop-in KSampler replacement that speeds diffusion for image, edit and video models without retraining.
A Hugging Face GGUF of DeepSeek-V4.1-Flash-1017 compressed to about 1 bit per parameter, with half its 384 experts per layer removed, yielding 72.1 GB of weights for code and agent workflows.
A dataset of 4,146 Three.js 3D scene examples pairing prompts with full thinking traces and complete HTML/JavaScript output, spanning objects, interiors, architecture, landscapes and voxel environments. Scene code was generated with MiniMax M3.1 Flash.
A Kling research paper argues video generation's bottleneck is shifting from rendering to semantic control, summarizing multimodal understanding insights from successive Kling model generations.
YANAMI is a large-scale multimodal dataset and benchmark for multiturn stateful interaction, aimed at improving unified multimodal models that must track information across turns.
A new paper proposes ToolSpace, a method that synthesizes realistic tool environments for training long-horizon agents to compose MCP services into executable workflows.
A research paper proposes orchestrating a heterogeneous ecosystem of specialized models, tools and policies into a unified AI system, and studies model and policy routing for general-purpose AI.
A research paper, DreamX, uses a video diffusion model to learn action representations that stay consistent across executions of the same action but invariant to embodiment and appearance, aiming for better transfer across robots.
A paper presents VidUEU-Agent, an automated agent pipeline that synthesizes instruction data from raw video for multimodal understanding, editing and unified tasks, aiming to cut annotation cost.
A research paper argues multimodal LLMs' compositional understanding is underused in composed image retrieval and proposes factorized MLLM embeddings to better map reference image plus text into a retrieval embedding.
A Hugging Face paper proposes local preference learning to compress large reasoning models' chain-of-thought, teaching them to stop deliberating once sufficient information is obtained and cutting inference costs.
A Hugging Face repository for ML-Intern-lab/Qwen3.5-0.8B-weight-space-programming, an image-text-to-text model produced via weight-space programming. It is a public repo, not an inference endpoint.
An open-source AI agent that brings Gemini directly into the terminal — v0.67.0 adds an IDE bridge and other changes. Free and open source for developers.
Capri is an open-source desktop AI agent with an Edgerunners-inspired look that works with Claude Code, Codex, Antigravity and other coding agents. The C# repo drew 138 stars in its first week.
Dify is an open-source collaborative workspace for building agentic workflows and RAG pipelines with support for many AI models and tools. Version 1.19.0 updates the self-hostable platform used to move prototypes to production.
A curated GitHub list of self-hosted open-source AI tools — local LLMs, chat UIs, agents, MCP servers, RAG and speech and image generation — health-checked weekly with dead projects removed automatically.
LiteLLM is an AI gateway that lets developers call 100+ LLM APIs in OpenAI or native format with cost tracking, guardrails and load balancing. v1.108.0 signs Docker images with cosign.
agent-autoloop is a TypeScript tool that has command-line agents take on and follow up tasks in the background, pinging you with cards only when they need a decision.
ComfyUI is a modular node-graph GUI, API and backend for running diffusion image and video models locally. Version 0.44.0 is out with updates to the open-source project.
A GitHub repo provides an uncensored AI API and browser chat built on DeepSeek and GLM models, with streaming plus Python and TypeScript examples, pricing and setup guides.
Open WebUI is a user-friendly self-hosted interface for running local and API models (Ollama, OpenAI). Version 0.13.0 adds a tabbed workspace for files, terminal sessions, artifacts, previews, voice calls and comments.
An open GitHub project that lets AI assistants such as ChatGPT and Claude remotely control a computer through MCP, with security controls. It drew 102 stars in its first week.
Balanov/skein is a Go project for orchestration of agent swarms that fails closed, using epoch-fenced work units and rejection of late completions. It drew 116 stars in its first week.
A Python repo of 32 Agent Skills for Claude Code and Codex aimed at Chinese doctors and medical students, covering topic selection, PubMed search, bilingual writing, journal submission and peer-review responses.
A TypeScript MCP server and CLI that finds Brazilian companies by CNAE code and region using open CNPJ data from Receita Federal.
Claude Code 2.1.317 is a terminal coding agent installed via npm that understands codebases, edits files and runs commands; Pydantic AI 2.65.0, Kiro 1.4.86 and Factory Droid 1.2.0 also shipped.
Raycast, a keyboard-driven launcher for controlling tools and running commands, released version 2.11.1.0 on macOS, Windows and Linux.
Apple's updated support document says the Siri AI waitlist used to gate access at launch in iOS 27 is no longer in use, except during high-demand periods.
After major Republican election losses, users asked Grok whether Elon Musk deleted his 'Vote Red Today!' post; the xAI chatbot said yes, per Yahoo. An example of AI answering questions about its owner's posts.
ChatGPT can now generate animated GIFs from text prompts or uploaded photos using OpenAI's Images 2.6 model. Users can adjust animation speed, edit frames and share results.
AT&T's Connected Life home security system now integrates with Gemini for Home, adding Gemini-powered event summaries and video history search. AT&T calls it the first home security product to use Google Home Premium.
Manus for Mac is an AI agent app that automates workflows on your own computer. Version 2.1.1 is a maintenance update for macOS, Windows and Linux.
ZDNET reports new Safari developer beta release notes point to deeper Siri AI integration and more Safari customization, plus three popular browser features.
Netflix is rolling out a Discover feature that recommends tailored streaming selections to help users find something to watch. ZDNET explains how it works and how to try it.
BoltAI is an AI chat client that lets you use multiple models in one desktop app. Version 2.18.5 is a maintenance release for macOS, Windows and Linux.
At ENGAGE 2026, Smartsheet launched vsActual, a standalone AI-native planning and execution product, plus an 'execution graph' aimed at linking team tasks to organizational strategy.
Siri Recaps, Apple's AI feature that summarizes audio and notifications, was tested in watchOS 27.3. Reviewers called it useful but raised privacy concerns and noted room for improvement.
A personal trainer in Wales built an AI-powered app aimed at improving gym training for users. Further product details were not specified in the report.
Google Photos is preparing Ask Photos features that could extract event details from photos, such as posters, and add them to calendars, with risks of AI errors.
MyHeritage introduced GAIA, an AI agent that helps users research family history and ancestry records.
ChatGPT can now lock private chats with a PIN, though users should know a catch. Separately, researchers warn one ChatGPT link could have exposed private chats and hijacked accounts.
Microsoft's Copilot app for iPhone and iPad reached version 2.115.3, a maintenance update with general quality and performance improvements.
Speechify 3.23.0 is out for macOS, Windows and Linux. It is an AI-powered reading and voice assistant app.
Tesla's software update 2026.38.6 reaches wide release, bringing the Grok AI assistant out of beta and adding Latvian language support.
Prelulu is an app for immersive stories with AI characters you can text or video roleplay with. It brings real-time avatar technology to consumer scale, and users can create their own characters with live avatar capability.
Travel platform Klook introduced an AI agent aimed at streamlining operations for its merchant partners.
Version 1.24.0 of Chatbox, a desktop app for chatting with GPT-4 and GPT-3.5 through the OpenAI API, is out for macOS, Windows and Linux.
A fan upgraded classic GTA 2 with next-gen graphics using Claude, while a Redditor used AI to get Nintendo Switch games running natively on an Apple TV, both hobbyist experiments.
OpenAI is running free ChatGPT training sessions for caregivers in five cities, aimed at helping them use the chatbot in caregiving work.
The PC gaming mods repository Nexus Mods will require authors of new mods to declare whether AI was used in their work, opting for mandatory disclosure rather than an outright ban.
Eney is an AI assistant from MacPaw, the maker of CleanMyMac and Setapp. It runs on Mac and iPhone and plugs into Apple services like Mail, Calendar and Reminders.
An open-source AI agent that brings Gemini directly into the terminal — v0.67.0 adds an IDE bridge and other changes. Free and open source for developers.
Cognition's SWE-2 Flash coding model runs three times faster and at one-fifth the price of SWE-2, available in Devin Desktop and CLI for routine software-agent work.
MariaDB released Enterprise Platform 2027, generally available, combining transactional, analytics, in-memory and AI capabilities with agent memory, retrieval and response caching; peak throughput up 56%.
Oracle is pitching a shared semantic layer and new security and AI-driven testing tools as enterprises pull data back into its database for real-time agent access.
Zed is a fast collaborative code editor; version 1.27.0 ships for macOS, Windows and Linux with its multiplayer editing features.
AWS Machine Learning · 2h
#
AWS explains connecting decision models Jev and Strands Decider to Amazon Bedrock AgentCore Evaluations via code-based evaluators, to answer narrow, repeatable agent-evaluation questions.
Baseten launched Grid, a public preview of a routing, gateway and guardrails layer for enterprise AI coding. Baseten says its Router cut cost per task nearly 70% with a 0.9-point quality tradeoff in initial benchmarks.
Contextbase stores AI agent context — semantics, knowledge, memory, traces and evaluations — inside PhoenixAI's database. It is in private preview; ObserveStack is the first app built on it.
Phone Harness gives AI agents control of a real iPhone or a rented cloud phone. Founder Shawn Pana open-sourced the harness for personal devices and sells persistent cloud phones.
The new VS Code release lets developers run Claude, Codex and GitHub Copilot agents within the same agent session, per Windows Report.
Replay's QA tool surfaced more than 900 bugs across 40 promoted apps, according to CEO Brian Hackett, and traced them to root causes AI agents can act on.
Redis's Iris real-time context engine now offers six composable capabilities for AI agents — four generally available, two in beta — adding request routing, reusable procedures, observability and resilient messaging.
AWS Machine Learning · 4h
#
A framework for production AI agents covering architecture, prompts, file-based tools and verification agents, implemented in working code from Deep Insight, an open-source multi-agent data analysis solution.
Nvidia's NeMo tooling turns production evidence into prioritized issues and custom evaluations to improve prompts, tools, models, memory and harness code for AI agents.
Cohere's Embed 5 and Rerank 5 retrieval models are now on OpenRouter, callable with the same API key and credits as other models without separate SDKs, accounts or bills.
datasette-llm is a plugin that adds LLM features to Datasette; version 0.1a10 fixes a bug where available models were not limited to those in the configuration.
Leidos introduced HeadWay Dispatch, software for overseeing and controlling AI agents across different vendors, aimed at organizations running multi-vendor agent fleets.
Oumi's Frontier AI Engineer agent can now build AI agents for customers and retrain the models behind them, targeting agent failures that emerge after demos.
Atlas is a free, MIT-licensed tool that creates a hidden second desktop on a Mac where computer-use AI agents can click and type while the user keeps working undisturbed.
Snowflake launches autonomous AI agents paired with tighter enterprise data controls, letting organizations run agents on their data while keeping governance in place.
Unite.AI outlines agent evaluation beyond final answers: measuring outcomes, inspecting trajectories, repeated trials, safety boundaries and the cost of reliable work.
An open-source campaign called OccupySaaS aims to use AI to clone common paid services such as Auth0 and Salesforce, targeting pricey service-as-software subscriptions.
A post examines HarnessMM, a token-reduction technique, in DeepSeek coding tests, describing results and behavior.
Fetch.ai added a Google Gemini integration, which Cryptonews says could change how its AI agents work.
A GitHub post describes OpenAI's one-week device project, putting AI engineering through a factory-style test.