DOYT is a to-do tool that pins current tasks to the iPhone Dynamic Island and lock screen. It supports long-pressing to complete or move to Later, an optional focus timer from 30 minutes to 8 hours, a history calendar, and Siri and Shortcuts integration. Core features are permanently free; the Pro version is a one-time purchase with a 3-day credit-card-free trial, requiring no registration and containing no ads.
As of September 2026, over 138 million AI Agent payment transactions have been completed on Coinbase's Base chain via the x402 protocol; total chain-wide transaction volume reached 165 million with amounts exceeding USD 50 million. 87% of individual transactions ranged between 0.01 cents and 10 cents. A Chainalysis report indicated that US wallets dominate model inference payments (88.1%), while Chinese wallets lead in code auditing (90%) and financial intelligence (46.7%). AWS, Google, Visa, Stripe, OpenAI, Solana Foundation, and others have integrated or supported x402. The median time to first spend from a wallet decreased from 73 days to 18 minutes, and the proportion of spending within one hour after topping up rose to 67%.
ihciah released Tinect, a self-developed three-layer overlay network tool. It supports destination-based routing, egress selection, and node path separation management, allowing configuration of preferred/candidate paths as well as independent outbound and return paths. It features dynamic FEC and Fake TCP link enhancement capabilities. The project is open-source, supports Linux/macOS, uses local YAML configuration and manual key exchange, and is currently in the experimental stage.
Shopify announced that it is opening its catalog of over 1 billion products to all AI shopping agents. Third-party agents can directly assist users with product search, price comparison, and checkout operations without requiring merchants to negotiate individually or integrate APIs. Consumers can link agents to their Shop accounts for one-click checkout via Shop Pay; orders are automatically synced to the Shop platform, but each transaction still requires confirmation by the buyer. Previously, this capability was only available to a few partners such as Meta Muse.
The EU Customs Reform has officially come into effect. Starting July 1, 2026, a temporary tariff of €3 will be levied on inbound shipments from outside the EU with a value not exceeding €150 per item; starting November 1, 2026, an additional Union handling fee of €2 per item will be imposed; e-commerce operators involved in systematic or repeated violations may face fines of up to 6% of the value of goods imported into the EU during the previous 12 months.
The author temporarily developed a photo album system for a wedding, supporting QR code upload, mobile number registration, multi-image upload, large image compression, duplicate file detection, photo wall, likes, backend review, large-screen cloud carousel, and weighted lottery. Photos are uploaded directly to Alibaba Cloud OSS, with Vue on the frontend and Laravel on the backend. The project is open source for reuse in weddings, parties, and other small events.
Microsoft has launched the Work IQ knowledge compilation engine, which leverages SharePoint to automatically transform scattered enterprise information into a structured knowledge base. This enables AI agents such as Copilot to achieve a 4x increase in response speed and an approximately 30% reduction in token consumption without changing the underlying large model. The feature is available to Frontier Preview tenants, supports Teams channels and Microsoft 365 group sites, and is currently enabled only for users on usage-based billing.
After Meta launched its C-end personal agent Muse and topped the iOS free charts in the US and Canada, ByteDance accelerated its personal assistant project codenamed "Spell," Tencent released LightVela and pushed Yuanbao and "Xiaowei" toward evolution into personal agents, and Alibaba's Qwen announced it would soon release "Qwen Go!"
Miao Yi Miao is a cross-device remote screen casting tool released by an individual developer. It supports casting video links from platforms such as Bilibili, Tencent Video, Youku, Douyin, AcFun, and Sohu to children's device browsers for playback without algorithmic recommendation interference. Parents can control it via WeChat Mini Program or web interface; free members are limited to 12 casts per day, while subscribed members cost 5.8 CNY/month. The website is miaoyimiao.net.
seventuning77 developed and launched henkanya.com, a pure front-end website for restoring Japanese mojibake. It supports mutual conversion attempts between UTF-8, Shift_JIS, and EUC-JP encodings, capable of restoring strings like "縺薙ｓ縺ｫ縺｡縺ｯ" to "こんにちは". For characters already lost as "?", it guesses based on context and marks them with underscores, while also providing full-width/half-width conversion functionality. All processing is completed locally without uploading data; the code was primarily generated by Claude, with the author responsible for requirements and testing.
Craft Agents is a fully open-source agent tool that runs locally, supporting integration with Claude, GPT, Gemini, or local models; it can connect to real tools (MCP, APIs, local files), inline render code diffs, data tables, charts, and live previews; it supports converting results into persistent Pages, teaching reusable Skills, running Automations on schedule, and provides desktop-grade UX.
Kunlun Tech released Matrix-Game 4.0, claiming it is the world's first video large model game capable of running on a single consumer-grade graphics card (NVIDIA RTX 5090), supporting 480p resolution, 60 FPS real-time rendering, and end-to-end latency below 150 milliseconds; the project has been simultaneously open-sourced, with a complete closed loop covering code generation for game logic, structural screen constraints, and real-time rendering by the video model; open-source addresses include GitHub, Hugging Face, and the project website.
This paper questions the validity of using teammate reasoning accuracy as the certification standard for theory of mind in multi-agent teams, noting that it fails to account for differences in error costs and error correlations. Validation on tasks such as Hanabi and task allocation showed that accuracy metrics can incorrectly approve harmful deployments or reject beneficial ones. The paper proposes the Blind-Spot Gate method, which achieves more reliable deployment certification by modeling error costs, pairwise interactions, and empirically measured dependencies.
This paper explores the path for general-purpose coding agents to generate images via executable programs, proposing that they outperform traditional generative models in structural fidelity and precise visual constraint satisfaction, but are limited in perceptual richness; it validates zero-shot drawing capabilities on the AgentDrawBench benchmark, and finds that reconstructed visual code retains high information content in visual question answering, with accuracy exceeding that of original images in some scenarios; furthermore, it proposes and preliminarily validates the hypothesis of an 'internal visual thinking language', showing that agents maintain consistent scene representations across drawing languages and single-pixel queries.
This paper presents The Agentic Scientist, an end-to-end distributed scientific research pipeline that supports collaborative work between human researchers and coding agents or fully autonomous computational scientific research. The system includes stages such as task drafting, idea generation, planning, sandbox experimentation, result aggregation, baseline and ablation construction, post-hoc analysis, and report writing, driven by a meta-controller that enforces verification of each stage against evaluation criteria. The authors state that the development version has been used to submit a paper accepted by ALIFE 2026, and has achieved cross-domain autonomous scientific discovery on open machine learning problems. Ablation experiments indicate that reliable scientific autonomy depends not only on foundation model capabilities but also on the cross-agent orchestration layer.
The paper introduces FlowSpec, a machine-readable specification for training large language model-based voice agents to adhere to statutory dialogue flows. In Thai outbound calling scenarios, testing on 4 unseen deployment processes (including 3 debt collection and 1 non-debt process) showed that document-based compliance check pass rates increased from 67–71% to 93–100%, and scenario pass rates rose from 15/43 to 40/43; audits identified 11 specification defects, of which 3 were not detected by its own metrics.
Archify v3.1 has been released, adding 5 new diagram types: ER diagrams, class diagrams, tree diagrams, timelines, and waterfall charts. Combined with the existing architecture diagrams, flowcharts, sequence diagrams, data flow diagrams, and lifecycle diagrams, it now supports a total of 10 diagram types. It supports generating interactive HTML diagrams from natural language and hand-drawn sketches, featuring zooming, detail viewing, and conversational editing capabilities. The project is maintained by an independent developer, and its GitHub repository has received 83,000 stars.
plane has released the open-source project harness-projects (https://github.com/chenhg5/harness-projects), aiming to provide a lightweight, localized, CLI-driven multi-agent collaboration solution for Claude Code and Codex projects. It supports shared context, reuse of skill tools, and automatic agent switching during failures. It does not rely on UIs and is designed specifically for native agent usage.
OpenNavo is a fully open-source graphical app store for Homebrew on macOS, supporting software discovery, installation, updates, uninstallation, and multi-language interfaces. The client is developed using Tauri 2 + Vue 3 + Rust; the project adopts the Apache-2.0 license, with all code hosted on GitHub, and introduces a FastMCP interface allowing AI Agents to participate in maintaining software information and multi-language content.
Adven released the LegionFanControl tool, providing one-click cooling and dust removal functions for Lenovo Legion Y-series laptops. It calls the firmware to push dual fans to full speed with a 120-second countdown, then automatically returns control to the EC; the interface displays real-time dual fan speeds, CPU/GPU temperatures, and the current smart fan mode, with status supporting automatic correction; based on Java 25 + JavaFX 27, it reads WMI via PowerShell, packaged as a GraalVM native exe or a JVM version dependent on JDK 25, requiring administrator privileges; only supports certain Y-series models; the open-source repository is acnxo/LegionFanControl.