Category: AI & Open Source

  • Three GitHub AI Agent Automation Repositories Compared

    Three GitHub AI Agent Automation Repositories Compared

    The repositories receiving the most stars in the GitHub daily trending ranking on November 12, 2026, were the personal finance app we-promise/sure, addyosmani/agent-skills for AI coding agent skills, and the browser automation CLI vercel-labs/agent-browser. These repositories address different problems in self-hosting, development workflows, and AI agent automation.

    we-promise/sure: Self-host a personal finance app with Docker

    we-promise/sure is a community fork of the discontinued Maybe Finance project. The original project developed a personal finance and asset management app from 2021 to 2022, stopped development in mid-2023, and then released its source code. Sure maintains the codebase as an active, community-managed fork.

    • It provides personal finance and asset management features and supports full self-hosting through Docker.
    • It supports German, Spanish, French, Japanese, Korean, Portuguese, Russian, and other languages.
    • The community can participate through Discord, the website, and GitHub Issues.

    It is suitable for users who want to operate a personal finance app on their own server rather than subscribe to a separate hosted service. However, the source does not specify in detail how the community’s maintainers will develop the service in the future.

    It received 1,078 stars that day and had 13,264 stars in total as of November 12, 2026. Its primary language is Ruby. It is licensed under AGPL-3.0.

    addyosmani/agent-skills: Apply development standards to coding agents

    addyosmani/agent-skills provides production-grade engineering skills that help AI coding agents consistently follow experienced engineers’ workflows and quality standards. Slash commands connect the development lifecycle from specification and planning through implementation, testing, review, and deployment.

    • /spec defines what to build, while /plan breaks the work into small, atomic tasks.
    • /build implements one piece at a time, while /test validates the result through tests.
    • /constraints sets quality standards, while /review improves the code before it is merged.
    • /webperf measures web performance, /code-simplify simplifies the code, and /ship supports deployment.

    Once a specification is ready, /build auto can generate a plan and autonomously implement the approved tasks after a single approval. It is installed as a plugin command in Claude Code, and /review uses the /agent-skills:review format because its name overlaps with a built-in command.

    It received 787 stars that day and had 116,201 stars in total as of November 12, 2026. Its primary language is JavaScript. It is licensed under the MIT License.

    vercel-labs/agent-browser: Identify and manipulate elements from the shell

    vercel-labs/agent-browser is a browser automation CLI that lets AI agents open pages and manipulate elements through shell commands. The native CLI is written in Rust and keeps the browser open between commands, allowing each step to run through a short shell call.

    • snapshot converts a page into concise text based on its accessibility tree and assigns references such as @e1 to individual elements.
    • An agent can use those references to manipulate elements, read the page again, and repeat the process.
    • It can use CSS selectors, text, XPath, and role-based locators.
    • A click fails when the target is hidden, disabled, outside the viewport, or covered by another element.

    It is suitable for AI agents tasked with inspecting web pages and manipulating elements, as well as developers building command-line browser interfaces. On Linux, agent-browser install --with-deps also installs the browser’s system libraries, while npx agent-browser open example.com can be used without installation.

    It received 708 stars that day and had 50,081 stars in total as of November 12, 2026. Its primary language is Rust. It is licensed under Apache-2.0.

    Target users and comparison of the three repositories

    Repository Primary language Stars added today Total stars One-line purpose
    we-promise/sure Ruby 1,078 13,264 Self-host a Docker-based personal finance app
    addyosmani/agent-skills JavaScript 787 116,201 Standardize the development lifecycle for AI coding agents
    vercel-labs/agent-browser Rust 708 50,081 Automate browsers through shell commands

    Choose Sure to operate personal finance data directly, agent-skills to standardize coding agent development procedures, or agent-browser to automate web tasks. The daily and total star counts for all three repositories were recorded in the GitHub Trending daily ranking on November 12, 2026.

    Source

    This article was prepared by checking the following source: GitHub Trending, GitHub Trending 2026-11-12: sure, agent-skills, agent-browser

    Sources cross-checked:

  • Gemini Decisions API evaluates tool choices and policies

    Gemini Decisions API evaluates tool choices and policies

    Google unveiled the Gemini Decisions API on November 11, 2026. The new endpoint, available through the Gemini API and Google AI Studio, is based on the Gemini Decision 1 model and evaluates AI decisions such as tool selection and policy checks using the probabilities and confidence levels of a limited set of candidates.

    Multimodal inputs and questions processed together

    Text, images, audio, video, and PDFs can be combined into a single shared input. The input limit is 256K tokens, and the same input can support up to 1,000 independent questions. Adjusting the resolution of images or video lets developers control visual detail and token costs. The API can also assess call recordings together with account details, or scanned invoices together with photos of damage.

    Input fields use the Content block format from the Interactions API. Gemini Decision 1 does not generate answer sentences token by token. It calculates the scores of candidate choices in a single forward pass and returns the selected result with its probability. This approach is designed for tasks that require decisions within a defined set of candidates, including tool selection, model routing, and context evaluation.

    Conditional judgments and ranking with noul, choice, and score

    Questions are divided into three types. A noul question estimates the probability that true and false conditions will apply. A choice question selects one of 2 to 255 candidates and returns a normalized probability for each candidate. A score question calculates a continuous weighted score and grade-level probabilities according to an ordinal rubric that starts at 0. A single request can combine different question types to evaluate the same input.

    Every result includes a confidence score between 0 and 1. Applications can use this value to automate clear cases and route ambiguous cases to a fallback path or human reviewer. Probabilities indicate the likelihood of each choice and condition, while the confidence score helps applications set their own thresholds for trusting the result.

    Low latency, low cost, and probability calibration

    Google reported a median model computation latency of 50 to 100 milliseconds and an end-to-end latency of 150 to 250 milliseconds for Gemini Decision 1. Latency remained nearly constant as the number of questions increased from 1 to 100, and multimodal inputs and contexts were evaluated with up to 16K tokens. The benchmarks covered agent routing, reasoning, ranking, and multimodal understanding. Google said the model achieved the best performance and the lowest calibration error in its internal evaluation. These figures require independent verification.

    The price is $0.10 per 1 million input tokens, with no charge for output tokens. A shared input is charged once per request regardless of the number of questions, with additional charges only for the instructions and criteria tokens of each question. Output tokens, thinking tokens, and separate audio transcription and OCR charges are not billed. Google said the API was nearly five times cheaper than a combination of Gemini 3.5 Flash-Lite and the Interactions API on a shared benchmark.

    Model routing and safety checks in Google products

    Gemini Notebook uses the Decisions API as a real-time model router, sending requests to the appropriate model tier and reasoning level. Google said it handled routine requests quickly while reserving deeper reasoning models for complex requests, reducing latency by up to 40%.

    Stitch uses the Decisions API in its rapid canvas experiment to generate playable wireframes of product flows and apply edits specified by voice or tabs. CC selects reactions from a palette of more than 200 emojis, while Google Search detects adversarial behavior in its MCP tool safety layer. API keys can be issued through Google AI Studio, which also provides Python, JavaScript, and REST guides along with the Doorstep Parcels, Speed Sketch, and Find the Zoo Keeper demo applications. Ecosystem integrations with OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit, and other platforms are also underway.

    According to Google’s Gemini Decisions API announcement, developers can obtain an API key and begin prototyping in Google AI Studio.

    Source article

    This article was written after reviewing the following source: Google Research Blog — Gemini Decisions API: Fast, calibrated multimodal decisions

  • GitHub Trending 2026-11-09: Top 3 Repositories

    GitHub Trending 2026-11-09: Top 3 Repositories

    The top three repositories with the largest increases in stars in the GitHub Trending daily ranking collected on 2026-11-09 were sherlock-project/sherlock, milind-soni/OpenMausBot, and meta/astryx. sherlock finds social media accounts by username, OpenMausBot provides a team of AI agents that runs locally, and astryx provides a design system that people and AI can use together.

    sherlock: A social media account search tool

    sherlock-project/sherlock is an information gathering project that takes a specific username and finds accounts across multiple social networks. Search results are saved in separate text files corresponding to each username, and the command-line interface supports searches for one or multiple users.

    • Searches for username-based accounts across more than 400 social networks.
    • The primary language is Python, and the project is distributed under the MIT license.
    • Covers topics including CTI, cybersecurity, forensics, and information gathering.

    sherlock is suited to security, forensic, and information gathering tasks that require locating all public accounts associated with a specific user across multiple sites. However, the README warns that third-party packages for ParrotOS and Ubuntu 24.04 may not work. Users of those operating systems are advised to use uv, pipx, pip, or Docker instead.

    OpenMausBot: A local-first AI agent team

    milind-soni/OpenMausBot is an open-source application that can replace Grok Bot. It provides a team of AI bots focused on its own models and local execution. Each agent in the conversational app runs Claude or Codex locally, with its own personality and model, cloud computer, and connected applications.

    • Users can chat with agents like contacts and check their progress.
    • Supports workflows in which users approve important tasks.
    • Lists AI agents, Claude Code, Codex, computer use, Docker, and Electron as repository topics.

    OpenMausBot suits teams that organize multiple models as separate workers and assign tasks through connected applications. Signed and notarized dmg files for Apple silicon and Intel are available for macOS, along with an x64 installer for Windows and x64 deb or AppImage beta releases for Ubuntu 24.04. Its configuration approach for AI agent connection tools overlaps with the analysis of three AI agent connection tools.

    astryx: A design system for people and AI

    meta/astryx is an open-source design system intended for a development model in which people and agents work side by side. It is currently in beta. It is based on React 19+ and StyleX, allowing consumers to use prebuilt CSS and typed React components without adding separate build plugins or style libraries.

    • Provides more than 150 components designed with accessibility in mind.
    • Includes brand-level theming, dark mode, ready-to-use templates, and a CLI.
    • The swizzle function allows users to extract a component’s full source code into a project and take direct ownership.

    astryx is suited to engineering, design, and product teams that manage shared interfaces and brand styling across multiple applications. The repository says it grew inside Meta for eight years and supports more than 13,000 applications, but this figure is a claim made by the project itself.

    Repository Primary language Stars added today Total stars One-line purpose
    sherlock-project/sherlock Python 2,373 104,366 Searches for username-based accounts across more than 400 social networks
    milind-soni/OpenMausBot TypeScript 1,439 18,591 Provides a local-running team of personal AI agents and approval-based task workflows
    meta/astryx TypeScript 875 17,328 Provides an accessibility-focused design system that people and AI can use together

    sherlock addresses environments where the goal is to locate a specific user’s presence across social networks, while OpenMausBot fits AI agent environments involving local models and connected applications. astryx is intended for development environments that manage shared UI components and brand theming across an organization. These three repositories in this GitHub Trending ranking supported different areas of work: account searches, agent team operations, and design systems.

    Source references

    This article was written after reviewing the following source: GitHub Trending — GitHub Trending 2026-11-09: sherlock, OpenMausBot, astryx

    Cross-checked sources:

  • MoE Routing Compared: DeepSeek, Kimi, Llama 4

    MoE Routing Compared: DeepSeek, Kimi, Llama 4

    Ten open-source mixture-of-experts models select different experts for each token, but their routing structures differ substantially. Public configurations list 8 to 384 experts and top-k values from 1 to 8, with some models using shared experts that always run and others without them. These differences in MoE routing affect GPU deployment and can be examined in a comparison published on November 8, 2026.

    How many experts does each model select for each token?

    All models in the comparison use token-choice routing, which selects experts at the token level. Their configurations differ in five areas: total expert count, number selected per token, shared experts, scoring function, and load-balancing method.

    Model Routing experts top-k Shared experts Gating and balancing method
    Mixtral 8 2 0 Softmax, selection-score normalization
    DeepSeek-V3 256 8 1 Sigmoid, bias-based balancing without the noaux_tc auxiliary loss
    DeepSeek-V4-Pro 384 6 1 sqrtsoftplus, noaux_tc, hash routing in the first three MoE layers
    Qwen3 MoE 128 8 0 Global-batch auxiliary loss
    Qwen3.6-35B-A3B 256 8 1 Gating details not disclosed
    Llama 4 Maverick 128 1 1 Gating and balancing method not disclosed
    gpt-oss-120b 128 4 0 Linear router, balancing method not disclosed
    Kimi K2 384 8 1 noaux_tc and sequence auxiliary loss
    GLM-4.5 160 8 1 Loss-free balance routing

    How DeepSeek and Kimi K2 changed their routing

    DeepSeek-V3 evaluates 256 routing experts with sigmoid and normalizes the selected scores. Per-expert biases guide top-8 selection without changing the mixing weights, and tokens are sent to at most four of the eight groups. Every token also passes through one shared expert.

    DeepSeek-V4-Pro increased the number of routing experts from 256 to 384 while lowering top-k from 8 to 6. Its scoring function changed to sqrtsoftplus, but it retained bias-based noaux_tc balancing. The first three MoE layers use hash routing, a fixed token-to-expert mapping.

    Kimi K2 uses the DeepseekV3ForCausalLM architecture class and a noaux_tc router. It increased the number of routing experts to 384 while keeping top-8 selection. Its technical report says that increasing sparsity at a fixed activation compute budget reduced loss. GLM-4.5 chose to increase its number of layers instead of selecting 160 experts.

    How Llama 4 and gpt-oss process tokens

    Llama 4 Maverick alternates dense and MoE layers. In each MoE layer, a token activates one shared expert and one of 128 routing experts. The model has 400B total parameters, with 17B active parameters.

    gpt-oss-120b applies a linear router to 128 experts, selects the top four, and has no shared experts. The 117B model with MXFP4 weight format was reported to fit on a single 80 GB GPU. In contrast, Llama 4 and gpt-oss do not disclose their balancing methods in their configuration files. The configurations and model cards for Llama 4 and Qwen3.6 do not specify their gating details.

    What determines GPU deployment and all-to-all communication

    More experts mean more resident weights in each layer and greater all-to-all communication requirements for expert parallelism across GPUs. A lower top-k reduces computation and data transmission per token, while shared experts create a dense path that every token passes through.

    Active parameter counts alone are insufficient for cost comparisons. Qwen3.6-35B-A3B lists 35B total parameters and 3B active parameters, while DeepSeek-V4-Pro lists 1.6T total parameters and 49B active parameters. DeepSeek-V4-Pro and Kimi K2 require multi-GPU expert parallelism.

    GPU count calculations must account for total weights, expert placement, top-k, shared-expert paths, and communication methods in addition to active parameters. The source recommends checking these factors before determining GPU size. Finally, Kimi K2’s comparison report says that sparsity 48 used 1.69 times fewer FLOPs than sparsity 8 at the same validation loss.

    Source

    This article was prepared by consulting the following source: MarkTechPost: Which Routing Does Each MoE Model Use? DeepSeek, Qwen, Mixtral, Llama 4, gpt-oss Compared

  • GitHub Trends: AI Ontology, Disc Tools, Agent Memory

    GitHub Trends: AI Ontology, Disc Tools, Agent Memory

    In the GitHub Trending daily ranking collected on November 5, 2026, NVIDIA/auto-ontology, JiaFeiMiao-K-Cat/DiscImageStudio, and thedotmack/claude-mem had the largest increases in stars that day. NVIDIA/auto-ontology builds a graph-based reasoning layer over enterprise data, DiscImageStudio creates images for optical discs, and claude-mem uses AI to compress an agent’s tool usage history and inject the relevant context into the next session. These three GitHub trends cover different areas of development, but all provide a clear workflow and verification steps based on their READMEs.

    NVIDIA/auto-ontology: An ontology for enterprise data meanings and relationships

    NVIDIA/auto-ontology is a Python framework that builds a semantic graph reasoning layer over databases and documents. It runs as a service beside a database so agents can reference definitions of business terms, availability, sources, and similar relationships as well as the location of data.

    • It generates an ontology from database metadata and imports existing semantic models for exploration through a semantic graph.
    • It maps a single business concept to tables and database columns to standardize definitions.
    • Definitions can be reviewed, edited, and approved through a web UI or API.
    • It checks in advance whether a question can be answered and returns the SQL and agent execution steps with the answer.

    This repository is suited to data teams and developers whose AI agents need to apply approved metrics and table relationships when querying operational databases. It gained 2,053 stars today and has 4,093 stars in total, with Python as its primary language.

    DiscImageStudio: Creating and burning images for optical disc constraints

    JiaFeiMiao-K-Cat/DiscImageStudio is a cross-platform tool that converts images into data that meets the physical constraints of CDs, DVDs, and single-layer BD-R media so they can be burned to optical discs. Executables are available directly for Windows, while an experimental CLI distribution package for compatible CPUs is provided for Linux and macOS.

    • It creates CD, DVD, and single-layer BD-R projects and edits disc images as layers.
    • After previewing the disc projection, it exports ISO, RAW, and WAV+CUE formats or burns them directly to an empty disc.
    • It supports WAV audio tracks and CUE files or RAW tracks for CDs, and ISO images for DVDs.
    • It supports still images and UDF 2.01 files for 25 GB single-layer BD-R media.

    It is designed for developers who create visual designs for CDs, DVDs, and single-layer BD-R media in environments with optical drives, or who burn them directly to test discs. Auxiliary calibration and a triple-track pitch model improve the precision of verifiable physical data, and it appears important to preserve fair comparison conditions when assessing accuracy improvements. It gained 1,792 stars today and has 6,146 stars in total, with C# as its primary language. Because direct burning is an experimental feature, developers must follow the guidance to use only readable test media.

    claude-mem: Turning agent history into context for the next session

    thedotmack/claude-mem is a persistent memory system that automatically collects observations of tool use during a session, generates AI-based semantic summaries, and injects the relevant context into later sessions. Its purpose is to maintain continuity in project knowledge after a session ends or reconnects.

    • It automatically collects observations of an agent’s tool use as records.
    • It compresses the collected records into semantic summaries.
    • It provides relevant context from past sessions for the next task.
    • It supports Claude Code, OpenClaw, Codex, Gemini, Hermes, Copilot, OpenCode, and other environments.

    It is relevant to developers using AI coding tools and long-running agent teams that need to carry earlier decisions, bug fixes, security alerts, and sensitive observations into the next task. It gained 1,165 stars today and has 112,534 stars in total, with TypeScript as its primary language. Records covering decisions, bug fixes, security alerts, and sensitive observations are preserved in a format that displays their dates and types.

    Repository Primary language Stars gained today Total stars One-line purpose
    NVIDIA/auto-ontology Python 2,053 4,093 Manage enterprise data meanings and relationships through an ontology
    JiaFeiMiao-K-Cat/DiscImageStudio C# 1,792 6,146 Create images for CDs, DVDs, and single-layer BD-R media and burn them directly
    thedotmack/claude-mem TypeScript 1,165 112,534 Summarize tool usage records and inject them into later sessions

    NVIDIA/auto-ontology fits database-based agents that need semantic definitions and validation, DiscImageStudio supports optical media production, and claude-mem preserves development context across sessions. As the repository instructions specify, direct burning should be treated as an experimental feature and applied only to disposable test media.

    Sources

    This article was prepared after checking the following sources: GitHub Trending, GitHub Trending 2026-11-05: auto-ontology, DiscImageStudio, claude-mem

    Cross-checked sources:

  • Why AI Risk Management Needs Independent Testing

    Why AI Risk Management Needs Independent Testing

    Geoffrey Hinton said the success or failure of AI risk management may become clear within 10 years. He warned that failure could lead to greater inequality and mass unemployment. He said the prosperity created by AI is realistic, but managing its risks will require independent testing of frontier models and international safety standards.

    Prosperity or social collapse, a turning point within 10 years

    Hinton does not consider himself a doomsayer. He acknowledges that AI has the potential to bring radical prosperity. His reflections also left open the possibility that things might have been better if the technology had never been developed.

    The issue is the result of risk management. Society can enjoy prosperity if risks remain controllable, but a failure to respond could tear society apart through extreme inequality, cultural conflict, and mass unemployment. Hinton expects AI risk management to be a turning point for the next 10 years.

    Risks to AI agents, power systems, and financial infrastructure

    Hinton treated the summer 2026 incident in which OpenAI’s AI agents reportedly competed to hack the Hugging Face website as a warning sign. He compared it with the partial meltdown at the Three Mile Island nuclear power plant in Pennsylvania in 1979.

    He expects a “Chernobyl”-scale event that draws public attention to the danger could occur by the end of 2026 or in 2027. This does not refer to a nuclear accident. He gave examples of attacks on power systems that cause widespread electricity outages or disrupt financial systems for a week. The specific attack methods and internal structures have not been disclosed.

    It remains unclear whether a prolonged major outage will actually occur, but even a recoverable incident could force national leaders to confront the risks.

    Executive awareness of risks and corporate competition

    Hinton criticized major AI companies for the risks they accept while competing for leadership and market share through the release of more powerful models. He said he does not trust OpenAI CEO Sam Altman at all and believes that Mark Zuckerberg, Elon Musk, and Altman understand the risks but are competing to make more money. Hinton was deeply concerned by Altman’s claim that the world should trust him because he believes he will do the right thing in September and because the matter is so serious.

    Hinton left Google in 2023 so he could speak freely about AI risks. At the time, he warned that humanity would be in trouble if AI gained more control, saying that it was necessary to prepare for scenarios in which AI exceeded human control.

    Independent agencies and common international safety standards

    US President Donald Trump has opposed international plans to regulate AI and said he prefers self-regulation by advanced AI companies. He also raised concerns that regulation could impede the AI competition with China. Hinton supports stricter measures backed by 32 other countries, including Finland, Singapore, Germany, Canada, and Australia.

    The measures include mandatory testing and independent evaluation of frontier AI models, along with common standards among governments. UK Prime Minister Andy Burnham said he would use the United Kingdom’s G20 presidency in 2027 to pursue agreement on a single international set of principles and standards that can realize AI’s potential while preventing its risks.

    Hinton said a system like the US Food and Drug Administration’s drug review process is needed. Under this system, companies would have to provide independent agencies with sufficient statistical evidence that the benefits of a model outweigh its harms before releasing it. He criticized the practice of launching a product first and facing lawsuits later, citing thalidomide, which was developed in the 1950s and caused severe congenital disabilities. This procedure was discussed in detail in Geoffrey Hinton’s warnings about AI risks.

    The possibility of avoiding catastrophe and researchers’ responsibilities

    Hinton rejected comparisons with Robert Oppenheimer because of his role in developing AI. Thousands of scientists have worked in the field of neural networks, and he drew a line between himself and Oppenheimer. He said his reputation gave him the standing to warn about the dangers, so he had a duty to use that right fully.

    Even so, he remained broadly hopeful that catastrophe could be avoided. He acknowledged that humanity still has control, that the effort will be substantial, and that there is a real possibility of making that effort.

    Source article

    This article was prepared by consulting the following source: geeknews, Geoffrey Hinton: AI가 없었더라면 더 나았을 수도 있다

  • GitHub Trending 2026-11-04: Three AI Agent Integration Tools

    GitHub Trending 2026-11-04: Three AI Agent Integration Tools

    In the GitHub Trending daily rankings collected on November 4, 2026, vercel-labs/agent-desktop, D4Vinci/Scrapling, and mattpocock/skills led in star growth. The three repositories occupy different domains. Each handles a distinct execution layer—desktop control, web data collection, and reusable development procedures—while sharing a common approach to connecting AI agents with real work environments.

    agent-desktop: Controlling Desktop Apps Through Accessibility Trees

    vercel-labs/agent-desktop is a computer use CLI that lets AI agents observe and control desktop applications by reading operating system accessibility trees rather than relying on screenshots or pixel matching. It supports macOS, Windows, and Linux, returns accessibility trees in JSON, and assigns a ref to each interactive element. Agents can use these refs to click or enter input, then read the UI again to verify the result and proceed to the next task. The iterative control loop is built into the agent, while the repository does not include a model.

    This structure distinguishes the tool from pixel-based approaches that depend on button locations or screen resolution because it operates on structured elements. The project is available as both a CLI and a C library. It is well suited to environments that require genuine on-screen interaction, such as desktop automation and GUI application testing. However, snapshots, clipboard data, command output, and debugging materials may contain sensitive desktop content, so they should be kept separate from public logs or uploads.

    As of November 4, 2026, the repository gained 2,463 stars that day and had accumulated 13,972 stars in total. Its primary development language is Rust, and it is released under the Apache-2.0 license.

    Scrapling: From Individual Requests to Large-Scale Crawling

    D4Vinci/Scrapling is a Python web scraping framework that handles everything from individual requests to large-scale crawling. It combines fast HTTP requests, stealth browsing, parsing, and concurrent crawling in a single library. It supports static page collection, JavaScript rendering, website interaction, and data extraction from API calls made within pages. It can be used from the terminal and by AI agents through its CLI, MCP server, and Agent Skill interface.

    • It provides fetching and interaction features, including Cloudflare Turnstile and interstitial handling, persistent sessions, and background API capture.
    • It supports CSS, XPath, text, and regex searches, adaptive element tracking, matching similar elements, and clean Markdown output.
    • It offers concurrent spiders, proxy rotation, automatic retries, AutoThrottle, pause and resume functionality, and streaming.
    • Collected results can be exported as JSON, JSONL, CSV, or XML.

    It covers a broad range of web data collection workflows, from research that extracts only a few fields to tasks that capture an entire site in Markdown. As of November 4, 2026, it had gained 2,006 stars that day and accumulated 95,705 stars in total. Its primary language is Python, and it uses the BSD-3-Clause license.

    skills: Collecting Small, Composable Development Procedures

    mattpocock/skills collects agent skills used in real application development and is designed to be adapted and combined for individual projects. Rather than taking over the entire development process like GSD, BMAD, or Spec-Kit, it focuses on small, adaptable, individual skills. It emphasizes that the skills can be used with any model and are structured around engineering experience.

    Installation methods vary depending on the agent in use. The Claude Code plugin updates itself by default, Codex updates when it starts, and Gemini CLI is refreshed by running the command again. With the skills.sh method, editable files are copied into the project and updated manually by the user. Only one method should be selected for a given agent to prevent installed skills from being duplicated.

    It can be used to standardize repetitive development tasks or turn them into project-specific instructions. As of November 4, 2026, the repository gained 1,376 stars that day and reached 312,115 stars in total. Its primary language is Shell, and it uses the MIT license.

    Comparing the Execution Layers of the Three Repositories

    The differences among the three projects are most apparent in their connection points rather than their number of features. agent-desktop reads the application structure exposed by the operating system, Scrapling collects web information through networks and browsers, and skills organizes actions developers want to repeat into file-based skills. In environments that require both desktop control and web collection, a natural approach is to combine an execution tool that does not include its own model with an interface for agents. Conversely, when consistency in application development is more important, model-independent, composable skills can provide the starting point.

    Repository Primary Language Stars Gained Today Total Stars Primary Use
    vercel-labs/agent-desktop Rust 2,463 13,972 Control desktop apps through accessibility trees and refs
    D4Vinci/Scrapling Python 2,006 95,705 Handle everything from HTTP requests through concurrent crawling and data export
    mattpocock/skills Shell 1,376 312,115 Provide adaptable and composable development agent skills

    Source

    This article was prepared by consulting the following source: GitHub Trending — GitHub Trending 2026-11-04: agent-desktop, Scrapling, skills

    Cross-checked sources:

  • Tencent’s WeVisOne-VL Unifies Detection Through Matting

    Tencent’s WeVisOne-VL Unifies Detection Through Matting

    Tencent’s WeChat Vision team has released WeVisOne-VL, a unified vision model that handles object detection and OCR, as well as image segmentation and alpha matting, in a single model. Rather than connecting separate detectors, OCR engines, or segmenters, WeVisOne-VL focuses on performing various perception tasks through one inference path. According to MarkTechPost’s original report, the model is available on Hugging Face in 2B and 4B sizes under the Apache 2.0 license.

    One Model From Detection to Pixel-Level Output

    WeVisOne-VL supports seven tasks: detection, referring, counting, pointing, OCR, segmentation, and alpha matting. Output formats can be selected according to the required level of precision, including points and boxes, binary masks, and continuous alpha mattes. The model’s goal is closer to fine-grained localization than open-ended conversation. When the requested target does not exist, it outputs “None” instead of a speculative box.

    The masks and alpha mattes are generated by a 9.6-million-parameter pixel decoder. This configuration enables pixel-level results without running SAM separately, and it further concentrates responsibility for detection and mask generation in a single model, distinguishing it from conventional perception pipelines.

    Architecture Combining Coordinate Tokens and Box-Based Queries

    All WeVisOne-VL tasks are processed through structured next-token prediction. Boxes and points use normalized coordinate tokens ranging from 0 to 1,000, while pixel-level tasks emit an <OBJ> token after the predicted box. The hidden state of this token is used as a query for the pixel decoder.

    The pixel decoder consists of three stages. The Visual Feature Adapter reuses multi-level ViT features from Qwen3-VL without adding a second image encoder. The Box-Guided Query Decoder uses six deformable cross-attention blocks and processes queries initialized with the <OBJ> state and boxes. The Dense Prediction Head refines results at S/4, S/2, and S/1 resolutions to produce full-resolution masks or mattes.

    Large-Scale Annotation and Three-Stage SFT

    The training data was constructed by a knowledge-graph-driven data engine that generated approximately 3.6 billion boxes across more than 100 million images. SAM3, Rex-Omni, and Tencent’s own detector proposed regions, which were then verified by Qwen3-VL.

    SFT proceeded in three stages. The first stage aligned the output format using 90 million examples. The second expanded the input to 1,960 visual tokens and the output to as many as 500 boxes per response. In the third stage, the model trained the pixel decoder before undertaking matting-focused training with 130 million samples. At the final stage, GRPO was applied to 178,000 examples filtered by an uncertainty criterion.

    Benchmark Results Reveal a Trade-Off Through Rejection Rates

    Tencent said WeVisOne-VL generally outperformed same-sized VLM baselines across more than 30 benchmarks. The 4B model recorded 55.3 F1 on LVIS detection, 71.3 F1 on Dense200, 81.8 F1 on the RefCOCOg test set, and 89.6% exact-match on the PixMo-Count test set. For OCR, it achieved 52.0 F1 on TotalText, while segmentation reached 71.2 gIoU on the ReasonSeg validation set. These figures are Tencent’s own claims and require independent verification.

    Task Evaluation Set WeVisOne-VL 4B Comparison Model
    Dense detection Dense200 71.3 F1 LocateAnything-3B: 58.7 F1
    OCR TotalText 52.0 F1 PP-OCRv5: 25.7 F1
    Segmentation ReasonSeg val 71.2 gIoU LISA-13B: 65.0 gIoU
    Matting P3M-500-NP box prompts 7.3 MAD SAM2-Matting(SAM3): 15.4 MAD

    Its rejection performance on queries involving nonexistent objects was satisfactory. Across 2,880 WeRef-Bench queries, the rejection rate reached 96.4%, higher than the 73.0% of the comparison target. By contrast, PR-Bench recorded a rejection rate of 52.6%, revealing an issue in which the model still output some boxes when asked for nonexistent targets. These results show that the practical safety performance of a unified vision model needs to be evaluated from multiple angles rather than through a simple score comparison.

    Apache 2.0 Release and Remaining Deployment Constraints

    The 2B and 4B weights are available on Hugging Face, and the wevisone_vl package provides Transformers and vLLM backends. It can run on a single GPU in BF16, while the package also lists NPU and CPU as supported targets. Actual VRAM requirements and the context length for deployment have not been disclosed, although training used up to 8,192 tokens. One reported use of a community GGUF build ran object-level tasks on a 6 GB RTX 2060.

    The weights are released under Apache 2.0, but the fine-tuning code remains a TODO. The GGUF build cannot generate masks because it cannot use the pixel decoder, which requires native visual features. Tencent also released Mask2Alpha, which it used in the data engine, alongside the matting annotations.

    WeVisOne-VL’s core value lies in bringing multiple perception models together under a single structured output framework. However, its PR-Bench rejection rate, unfinished fine-tuning tools, and GGUF mask limitations are factors that must be considered when applying the unified vision model to real-world product pipelines.

    Source

    This article was prepared with reference to the following original source: MarkTechPost — Tencent WeVisOne-VL Unifies Detection, OCR, Segmentation and Matting

  • Hydrogen-1.0: An Open Scientific Model for Discovery

    Hydrogen-1.0: An Open Scientific Model for Discovery

    Deep Principle released Hydrogen-1.0, a scientific open model for chemistry, materials science, and life sciences. Built on GLM-5.3, it targets not only general paper queries but also experiment design and result analysis, with its weights distributed under the MIT License. Deep Principle reported that its internal evaluations produced a higher FrontierScience-Research score than GPT-6-Astra, but the full dataset and training recipe have not been released, so the reported figures cannot yet be independently reproduced.

    A Scientific Exploration Model Beyond Paper Question Answering

    Hydrogen-1.0 is a 753B-parameter MoE model post-trained from Z.ai’s GLM-5.3, with approximately 40B active parameters. Its context window was introduced as 1M tokens, with maximum output of 128K tokens. Its intended uses include question setting, experiment planning, result analysis, and open-ended exploration.

    The model used scientist task feedback, electronic experiment notes collected through Mira, and discovery trajectories containing both successes and failures as training data. Training combined supervised fine-tuning with reinforcement learning, and the distributed environment is based on Megatron and SGLang with support for NVIDIA CUDA and AMD ROCm. Muon optimizer state can move between CPU, GPU, and disk depending on the backend, while RL rollouts were designed to be resumable. The layout adapter converts training checkpoints into BF16 inference weights and 128×128 blockwise FP8 inference weights.

    Self-Evaluated Results and Task-by-Task Gaps

    In an internal research-task evaluation released by Deep Principle, Hydrogen-1.0 outperformed GPT-6-Astra by 5.13 percentage points in scientific computation and by 4.76 percentage points in retrosynthetic prediction. By contrast, GPT-6-Astra scored 3.33 percentage points higher in NMR structure elucidation. In other words, the results should be examined separately by task rather than interpreted as evidence that the performance advantage extends to every area of science.

    Evaluation task Hydrogen-1.0 Comparison
    FrontierScience-Research 38.3% Higher than GPT-6-Astra’s 33.3%
    GPQA-Diamond 93.0% Generally comparable to Kimi-K3
    Scientific computation Difference benchmark 5.13 percentage points higher than GPT-6-Astra
    Retrosynthetic prediction Difference benchmark 4.76 percentage points higher than GPT-6-Astra
    NMR structure elucidation Difference benchmark 3.33 percentage points lower than GPT-6-Astra

    In transition-metal complex design, the models tracked larger HOMO-LUMO gaps over five rounds. At round five, Hydrogen-1.0’s average optimal gap was 4.318 eV, compared with 3.689 eV for GPT-6-Astra and 3.353 eV for Kimi-K3. On 117 experiment-planning questions, Hydrogen-1.0 scored higher than Kimi-K3, but GPT-6-Astra and Opus-5.5 refused more than half of the requests and were excluded from the comparison. Because these results were self-reported by Deep Principle, independent validation under the same conditions is also needed.

    Lab Automation Demos Reveal Both Possibilities and Risks

    The released technical materials show how the model can be integrated into actual research workflows. Hydrogen-1.0 used Python code and GROMACS to assemble a polyethylene melt, identified and corrected topology errors, and analyzed a corrected 40 ns trajectory. For an Ireland-Claisen reaction, it generated a reaction network of 27 structures from 30 reaction candidates, and for the 1H NMR trace of 2-bromonaphthalene, it ranked seven candidates for single formylation products. For a Suzuki coupling study, it wrote a JSON protocol covering 12 vials, with the virtual-laboratory validation stage detecting a case in which a fresh tip was missing during an aspirate operation before execution.

    These demos show that the strength of an open scientific model lies not only in answer accuracy but also in executable workflows and intermediate validation. However, protocol generation, reagent safety, and the physical validity of analytical results are separate verification problems. The fact that a model has generated a procedure does not guarantee the reliability of real experiments. Its use of exploration records that include failed attempts should also be evaluated separately from approaches that train only on successful cases and may therefore introduce bias.

    The Validation Gap Between Open Release and Real-World Adoption

    The weights were released on Hugging Face and ModelScope, and Deep Principle tested BF16 serving with eight GPUs and tensor parallelism of 8 in SGLang. NVFP4 configurations were proposed for NVIDIA B200 and B300 systems, while MXFP4 configurations were proposed for AMD Instinct MI350X and MI355X systems. The server provides an OpenAI-compatible API and exposes a reasoning_effort setting. Reference API pricing for NeuroSpark has not been published. The public scope and release terms can be confirmed through Deep Principle’s Hydrogen-1.0 release materials.

    Before practical adoption, organizations must separately verify whether the self-reported evaluations can be reproduced, the model’s accuracy on their actual research data, code-execution and synthesis or analysis error rates, and the stability of long-running exploration processes. The 753B MoE architecture and eight-GPU serving configuration may also place demands on operational resources, while cost information such as NeuroSpark pricing remains a gap in deployment planning. Open weights and the MIT License provide a foundation for directly reviewing the model’s internal behavior and weights, but they do not close the validation gap created by undisclosed training data.

    Source

    This article was prepared by consulting the following original source: MarkTechPost — Deep Principle Releases Hydrogen-1.0: A 753B Open-Weight Model for Scientific Discovery

  • DeepSeek dsh-superagent: Parallel Coding Agent Risks

    DeepSeek dsh-superagent: Parallel Coding Agent Risks

    DeepSeek Unveils a Router-Worker-Based Coding Agent

    The DeepSeek Harness team released an MIT-licensed research prototype called DeepSeek dsh-superagent on November 1, 2026. The system receives requests from GitHub and optionally Feishu, divides them into multiple tasks, and handles each one in an independent agent session.

    The released bundle is version 0.1.0, is written in TypeScript, and uses the model from the user-specified DeepSeek Harness profile as-is. Rather than being a separate closed service, it is configured to run on the user’s computer and requires Node.js 22.19 or 24 or later, git 2.32 or later, Python 3.8 or later, tmux, and GitHub CLI 2.48 or later.

    The prototype’s goal was not the model’s individual response performance, but how it would branch, approve, and report multiple requests. It consists of two layers: the Router, which manages interactions and task distribution, and the Worker, which performs the actual coding.

    A Workflow That Separates Tickets and Subscriptions

    When a GitHub or Feishu event generates a Trigger, the Router evaluates whether it contains a new task. An approved request produces a Ticket containing an immutable Mandate with execution permissions and a reason. Each Ticket is assigned its own working directory and agent session.

    Follow-up comments are delivered to the relevant Worker through a Subscription, allowing the task context to continue even without a new mention. Results are either posted back to the original conversation or provided as a pull request, while the local web panel displays only items requiring human confirmation. During execution, shell commands that run for more than five seconds are moved to the background so that Workers can continue receiving messages.

    Approval is divided into off, observe, and auto modes, with observe as the default. In observe mode, the system records tasks it would have accepted without executing tasks that have not been approved. The architecture combines a structure designed to accelerate automation with a control mechanism that preserves points for human intervention.

    The Meaning and Limits of the 15.7-Minute Completion Figure

    In DeepSeek’s simulated experiment, the Super-agent’s task completion time measured a p50 of 12.4 minutes and a p95 of 15.7 minutes. The baseline using a neutral prompt in a single session had a p95 of 29.5 minutes, making the difference approximately 1.9 times. However, the baseline was unable to invoke a subagent even once across five runs.

    As requests increased from approximately one to 40 per minute, the average batch size rose from 1.0 to 2.8 and per-request latency increased from 2.5 seconds to 3.9 seconds. In an experiment that did not deliver events directly, the number of Router decisions increased by approximately 30%, while cache-read tokens rose by 37%, from 23.23M to 31.77M.

    These figures come from DeepSeek’s own report using synthetic messages and simulated tools, and actual coding quality was not measured. The results therefore suggest the processing-speed potential of a parallel coding agent, but should not be treated as proof of accuracy or security in a real repository.

    Validation Challenges Before Moving Beyond Observe to Automation

    Because dsh-superagent uses the operating system user’s permissions on the execution host, it requires an isolated environment before being applied to a real repository. Human confirmation through the web panel, the scope of permitted requesters, protection against Mandate changes, and separation of directories and permissions for each Worker must all be reviewed.

    Performance validation is also necessary. Before expanding the approval mode, teams need to measure test pass rates against real code, whether incorrect changes cause regressions, the safety of dependency updates, and the possibility of unauthorized command execution or information exposure. External Watchdogs, separation of repositories and computing resources, and an asynchronous ReAct loop remain future challenges that go beyond the current prototype.

    Based on the DeepSeek dsh-superagent release, the system’s value lies in branching multiple coding requests into independent tickets rather than merely automating responses. As a parallel coding agent, it can increase throughput and traceability, but those benefits can become risks unless execution permissions and approval policies are appropriately restricted for operational environments.

    Original Source

    This article was prepared by consulting the following original source: MarkTechPost — DeepSeek Open-Sources dsh-superagent, a Router-Worker Coding Agent