Juejin’s first installment in an Agent development series explains how large language models differ from systems that can use tools, maintain state, and execute multi-step tasks. Using an in-app-purchase customer-support scenario, it connects Function Calling, Runtime loops, ReAct, Workflow, Planner-Executor, Multi-Agent patterns, Context Engineering, RAG, evaluation, observability, model gateways, security, approvals, and recovery into one engineering framework. The article also previews 12 follow-up lessons that build a runnable teaching project around order lookup and coin-credit troubleshooting.
量子位 reports that SpaceXSI chief engineer Lauren Tan used her AI Engineer CODE talk to outline how the Grok Bot team manages AI-assisted development. Her workflow prioritizes clean code, lint rules, architectural constraints, runtime verification and formal verification over natural-language guidance, while using tools such as TLA+, Lean and Bombadil to check agent-generated work. Tan also described an automated release-testing loop in which agents explore applications, create fix pull requests and may merge them; her PR-volume figures were not independently verified by the report.
The mjmirza/app-store-compliance repository provides a playbook and pre-submission guard for iOS and Android apps, including checks for privacy manifests and Data safety forms. Its author says the project documents 117 rejection patterns and fixes and tracks 95 regulatory deadlines. The guard can run through AI coding tools and block submission commands when it detects critical risks. The repository uses the OpenRoots Agent License 2.3 for its current release, and earlier releases have different license terms.
Baseten says it secures more than 100 inference clusters across more than 20 cloud providers, including hyperscalers and neoclouds. Its security approach includes per-cluster hardening, network and tenant isolation, scoped credentials, isolated build environments, runtime detection, centralized logging, and an accompanying checklist for evaluating security standards.
CNBC, citing a House report, says SpaceX subsidiary Cursor told lawmakers it is considering phasing out Chinese-developed open-weight AI models because of geopolitical risk. The report does not specify which models would be affected, implementation details or a timeline.
Valhalla Matrix Governance Laboratory reviewed Google Labs’ stitch-skills as a collection of standardized Agent Skill contracts for exporting designs, generating components, and validating artifacts through Stitch MCP Server. The repository separates React and React Native workflows, includes SWC-based checks for color literals, and adds timeout and input validation to static HTML extraction. The review was limited to static source analysis and did not test runtime compatibility, service connectivity, or generated-output fidelity.
钛媒体 and 36氪人工智能 published the same analysis of AI product closures, including Sora, 冒泡鸭, XEVA, 字节扣子企业版, 腾讯元器, 百度梯子AI and 米哈游 AnuNeko. The reports say Dang.ai had marked 1,895 discontinued tools and services by November 20, 2026, while Sensor Tower counted 3.8 billion generative-AI mobile-app downloads in 2025 and more than $5 billion in in-app purchase revenue. The analysis attributes closures to pressure from rising inference costs, feature convergence, limited differentiation and changing product portfolios, rather than to an overall collapse in demand.
The author outlines Morde P0, a proposed minimal workflow in which one Agent handles one repository-scoped bug or request and returns a pull request for human review. The design separates execution from delivery decisions, requires evidence tied to the candidate commit and runtime environment, and uses persistent state to prevent duplicate external actions after interruptions. The author also warns that the described Y Combinator case does not show that teams universally skip code reading or human review.
The article presents Agent Skills as folder-based packages for reusable instructions, scripts, references, and assets. It explains a progressive loading model: an agent initially sees each skill’s name and description, reads the SKILL.md body when the description matches a task, and accesses scripts, references, or assets only as needed. It also distinguishes knowledge skills from workflow skills, compares them with global rules and short prompts, and describes manual, command-line, and agent-assisted installation, automatic or manual invocation, and cautions against installing vague or unnecessary skills. The report attributes the open specification’s origin to Anthropic and says it was later supported by tools including Codex, Cursor, and DeepSeek Harness, but the supplied report does not provide independent confirmation of that history or compatibility.
In a Latent Space interview with Linear COO Cristina Cordova, the company said agents created 3% of Linear’s work a year ago, while a later measurement said more than half of the issues created in Linear came from agents. The report did not clarify whether the two figures use identical definitions or measurement periods. It also said Linear has shipped an MCP server, a platform for building agents into Linear, integrations with Codex, Cursor and GitHub Copilot, and its own Linear Agent. The report stated that Linear passed $100 million in annual recurring revenue this year, with more than 40,000 paying customers and 177% net revenue retention; these figures were not independently verified.
A tutorial by Xiong Yun reports that Alibaba Cloud’s AgentLoop plugin is available in the Qoder plugin marketplace. It describes natural-language workflows for generating a LoongSuite Pilot enterprise installer, querying agent usage and Token consumption, and reviewing risk events and user intent. Administrators can also turn query results into dashboards with filters and drill-downs, preview them locally, and publish them to Alibaba Cloud SLS after confirmation. The tutorial also says the installer can be distributed through Qoder’s enterprise marketplace and install the collector when a member’s Qoder session starts.
Appwrite's project repository describes the platform as an open-source, agent-first environment for building applications, with a hosted MCP server that lets coding agents operate on live projects through OAuth. It also supports managed cloud and self-hosted deployment, alongside Auth, Databases, Storage, Functions, Messaging, Realtime, and Hosting. These capabilities and integration details come from Appwrite's own documentation and have not been independently verified.
Xiaohu tested HeyGen-associated open-source video framework HyperFrames with WorkBuddy and the free-tier Hy4 flash model, producing three videos in one night. The workflow generates a webpage and records it frame by frame, with storyboard planning, self-review, render interruption for problematic elements, and post-render checks; a 15-second animation took about ten minutes, while a 203-second talking-head video rendered in 1 minute 58 seconds. The author says HyperFrames is suited to structured product videos, data animations, and subtitled talking-head clips, but requires coding and animation knowledge and is silent by default.
Google DeepMind released Gemini 4 Neon, the fourth Gemini 4 model in roughly seven weeks, with a 1M-token context window and four effort levels. Latent Space reports pricing of $3 per million input tokens, $0.15 for cached input and $15 for output, with rollout planned for the paid Gemini API, AI Studio, Android Studio, Gemini for business and Antigravity. Preliminary results placed Neon second on Agent Arena and the Vals Index, while Artificial Analysis reported that it matched Carbon’s Intelligence Index score of 60 at high effort.
An author presents several options for users whose Claude accounts have been blocked: overseas Mac mini or cloud infrastructure, multiple accounts, overseas VPS setups, alternative coding agents, and a bundle of domestic agent services. The author also recommends using Magpie to route domestic providers through DeepSeek Harness or Claude Code and turning Claude workflows into Skills for domestic models. The cited prices, quotas, account-opening process, routing behavior and network isolation are the author's claims; the report provides no independent verification.
Cursor announced that Gemini 4 Neon is now available in its editor. Cursor says the model scored 63.7% on CursorBench 4.1 at its default effort, compared with 57.3% for Gemini 4 Hydrogen, and priced it at $3 per million input tokens and $15 per million output tokens. The benchmark comparison is Cursor's own claim and has not been independently reproduced in the supplied report.
Snapchat launched Easy Lens CLI, a command-line tool that lets developers create AR Lenses through AI coding assistants such as Claude Code, Cursor, and Codex. The tool supports Mac, Windows, and Linux, and Snapchat says it can use more than 100 AR building blocks and 50 AI asset generators to generate assets, preview and test Lenses, publish updates, and track performance; these automation and capability claims have not been independently measured.
LangChain says LangSmith for Coding Agents is now available in public beta on LangSmith Cloud, linking agent traces and spend to GitHub pull requests and GitLab merge requests. A new overview dashboard aggregates costs by pull request, work type, model, and pull request state, while tracking work that does not produce a pull request as unattributed spend. The release supports Claude Code, OpenAI Codex, Cursor, GitHub Copilot in VS Code, and Deep Agents.
Modem Blog says Modem 2.0 is now available, replacing the company’s earlier standalone AI product-manager approach with a “customer product graph.” The graph groups customer reports from channels such as chat, support tickets and sales calls, then links each topic to its status, people, tickets, pull requests and code changes. Modem says users can access it through Claude Code, ChatGPT, Cursor, Devin or other MCP clients, while its own agent and configurable automations can create issues, track fixes and prepare customer follow-ups.
A Juejin author compares TRAE, GitHub Copilot, Cursor, Windsurf, CodeBuddy, and Tongyi Lingma across pricing, feature completeness, Chinese-language adaptation, and related criteria. The ratings combine the author's personal testing with cited community evaluations, and the author explicitly notes that they do not represent an industry-wide conclusion.
A Juejin article proposes combining mid-run cognitive handoffs, end-of-run cognitive settlement, and multi-session continuation to preserve continuity during long AI tasks. It argues that agents should persist structured state between shorter sessions rather than rely on one session lasting hundreds of turns. The article also makes unverified claims about token savings, context degradation thresholds, and comparisons with Claude Code and Cursor.
SpaceXSI says it has built five Grok Bots that independently own scoped accounting processes, including customer provisioning, billing, payment application and collections. The bots escalate exceptions and seek approval for sensitive actions such as credits and refunds. SpaceXSI says a typical new bot now takes about three weeks to onboard, while the first bots took longer: Billing Support Bot took four to five weeks and Collections Bot about three months.
孟健 tested the MIT-licensed T3 Code by using a phone to direct Claude Opus 5.5 while it summarized several documents into a structured summary.md weekly report. The author reports that T3 Connect can monitor progress, send instructions, synchronize changes to the computer, and support multiple agents and parallel worktrees, though these details come from a single user's account.
Juejin AI author Lin Runqi (Taiye) introduces a workflow combining LoongSuite Pilot with a Qoder installer plugin and Qoder Enterprise Marketplace for organization-wide AI coding telemetry. Pilot reportedly normalizes activity from 20 coding agents into a shared GenAI event schema, while the installer distributes and maintains the collector with configurable collection scope and output redaction. These capability and coverage claims come from the article author and have not been independently verified.
Reverse engineering of MissionControlAggregator.dll found that Windows 11 records bitmap indicators showing which coding agents and developer tools were observed, including Codex, Claude, GitHub Copilot, Cursor, Python and .NET. The researcher verified on a US-configured Windows 11 Home ARM64 installation that the report entered DiagTrack's upload queue when optional diagnostic data was enabled, which was the default in that setup. The report contained no prompt text or source code, and its timestamps represented collection windows rather than per-tool usage duration. Behavior in other regions, editions or configurations was not established.
OpenCode Browse is an MCP server and Chromium extension that lets AI agents operate a user's real browser, including existing tabs, logins and extensions. According to the OpenCode Browse product page, agents can open background tabs, read pages, click, type, capture screenshots, record tabs and monitor network traffic, while handing login, two-factor authentication and payment steps back to the user. The tool is advertised as compatible with MCP clients including OpenCode, Claude Code, Cursor, Codex and Zed.
Google has made AdMob agent skills available for coding assistants such as Claude Code, Codex and Cursor, allowing them to add banner or full-screen ads, migrate ad code and check setup errors. The Gemini-powered Ask AdMob assistant and an MCP server for account management remain in testing, while additional agent actions requiring user approval are planned for 2027.
Island found 186 typo-squatted domains impersonating 27 popular AI brands, including ChatGPT, Claude, Perplexity, Midjourney, Cursor and DeepSeek. The sites redirected visitors toward ClickFix-style commands, a browser-hijacking application called Net Guard, or malicious extensions that altered search settings and harvested queries. Researchers said the campaign remained active and linked it to 211 entry domains, including lookalikes for banks, insurers, telecom firms and enterprise software providers.
Satya Nadella said in The Pragmatic Engineer interview that Claude Code caught Microsoft off guard by moving coding toward the command line, while Cursor’s agent mode prompted Microsoft to respond. He argued that as AI-generated code becomes cheaper, verification becomes the harder part of software development, and described Microsoft’s “software factories,” slower junior hiring, changing team structures and voluntary exit program. Nadella also said Microsoft and Carnegie Mellon University launched the PRAISE mentoring program for AI-assisted software engineering.
36氪, citing 壹娱观察, reports that Artificial Analysis scored Fable 5.2 at 65 in its composite intelligence index, placing it first. In an Anthropic demonstration, 1,000 agents allegedly completed a formal proof of Fermat's Last Theorem in 30 hours, while Opus 5.5 reached 27% of the same work in the same period. The analysis also notes that Fable 5.2 cost an average of $26.54 per task at its highest tier, compared with $10.04 for Opus 5.5, and that it lost to Gemini 4 Carbon in an office-automation test.
What problem it solves
Web teams and developers can need to rebuild an existing site in a modern codebase when migrating platforms, recovering lost source code, or learning how a production site is constructed. This template gives an AI coding agent a repeatable workflow for recreating a site as an editable Next.js application while comparing its appearance and behavior with the source.
Key featur…
What problem it solves
AI agents often produce plausible SQL without the business definitions, approved joins, units, and examples needed to make results trustworthy. Teams also need a reviewable way to manage that context and turn governed answers into shareable dashboards across existing data sources.
Key features
Governed semantic and context layers — Business meaning is represented through the…
What problem it solves
Personal and team knowledge is distributed across notes, conversations, trackers, and research, making it difficult to capture, organize, verify, and synthesize useful information. COG provides a markdown-based workflow in which agents can maintain a shared knowledge base, support recurring research and operational tasks, and verify selected work through independent checks.
…
What problem it solves
AI agents need product-specific instructions to use NVIDIA software reliably across physical AI, robotics, simulation, CUDA-X libraries, RAG, and platform workflows. The catalog centralizes these skills while keeping them maintained in their respective product repositories and synchronized through an automated pipeline.
Key features
Agent-discoverable skill catalog — The NVI…
What problem it solves
Developers and coding agents need to understand how code is connected before making changes. GitNexus provides a graph-based view of symbols, callers, dependencies and execution flows to support code discovery, impact analysis and refactoring.
Key features
Graph-based code intelligence — Index a codebase as a knowledge graph covering symbols, calls, dependencies and executio…
What problem it solves
Agent memory is often opaque: stored context cannot be easily inspected or understood. OpenViking addresses this by organizing knowledge, memory, and skills in a browsable virtual filesystem, with summaries that help agents identify relevant context before reading full files.
Key features
Unified context filesystem — Knowledge, memory, and skills share one filesystem address…
What problem it solves
Traditional development accumulates technical debt and leaves important knowledge scattered, forcing later changes to rediscover context. The plugin addresses this by turning each engineering cycle into a structured workflow and writing durable learnings into artifacts that subsequent cycles can use as grounding.
Key features
Structured engineering loop — A six-step cycle mo…
What problem it solves
The hidden system instructions that shape AI product behavior are difficult to inspect and change across models, products, and versions. This repository makes collected prompts available in a structured, product-oriented archive for study and comparison.
Key features
Broad product coverage — The archive covers prompts from major AI providers and products, including Anthropic…
What problem it solves
Chrome DevTools for agents (`chrome-devtools-mcp`) lets your coding agent (such as Antigravity, Claude, Cursor or Copilot) control and inspect a live Chrome browser. Both give your AI coding assistant access to the full power of Chrome DevTools for reliable automation, in-depth debugging, and performance analysis.
Key features
Browser control and inspection — Give coding age…
What problem it solves
AI agents need relevant design context from Figma files to generate code that follows a design system and reflects design intent. This project documents how to connect the Figma MCP server to supported clients, use its tools and skills, and apply practices for more consistent results.
Key features
Connect supported MCP clients — Setup instructions cover the Figma MCP server …
What problem it solves
Coding agents need a structured way to create, manage, and build from Google Stitch designs while integrating with the Stitch MCP server. The project organizes design, code generation, and supporting workflows into skills that can be installed with supported coding agents.
Key features
Design workflows — The design plugin covers converting frontend code into Stitch designs, …
What problem it solves
AI coding agents lose context between sessions, so the same decisions get re-explained and nothing about the project is remembered. They also lack a shared workflow and evidence, leaving the human to read every line to confirm what was done. This project adds memory, a workflow, and evidence to agents you already have.
Key features
Persistent project memory with Engram — Eng…
What problem it solves
Developers need a consistent way to control locally configured coding agents across mobile, web, and desktop environments. T3 Code addresses this by providing multiple client applications, support for several agent providers, and remote access to a local server.
Key features
Control from mobile, web, and desktop — Use the mobile, web, or Electron-based desktop application to…
What problem it solves
Teams that rely on AI agents need a way to give each agent a role, workspace, and instructions, keep context across conversations, and control what agents are allowed to do. Without that, agents run blind: actions happen without review, memory does not persist, and repeating scheduled work means redoing setup by hand.
Key features
Create agents and teams — Give an agent a na…
What problem it solves
Coding tools use different component formats, model names, permissions, and discovery rules. Contributors need a shared way to write reusable plugins, agents, skills, and commands once while producing compatible artifacts for each supported harness.
Key features
Multi-harness component support — One collection of plugins, agents, skills, and commands can be used across Claud…
What problem it solves
AI coding assistants can generate interfaces with visual and structural patterns that feel machine-made, such as repetitive headings, empty sections, misplaced heroes, and forms that do not complete their intended action. Hallmark applies design constraints, product-aware page structures, browser-based linting, and fresh-context review to help generated interfaces look desig…
What problem it solves
Developers need a coding agent that can work directly from their local development environment without requiring a separate desktop application. The project provides installation paths for terminals, IDEs, and desktop use, along with authentication through a ChatGPT plan or an API key.
Key features
Local terminal coding agent — Run Codex CLI locally on your computer to use a…
What problem it solves
Agent workflows, corrections, debugging patterns, and repository conventions often disappear when a session ends or remain trapped inside one coding tool. Beacon addresses this by capturing session history across harnesses and turning useful findings into reusable memory for future agents.
Key features
Cross-harness session history — Capture session history from Claude Code,…
What problem it solves
DeepSeek Harness workspaces, tools, and project setup normally remain tied to the computer running the Harness instance. The project addresses the need to continue using that environment from another computer, phone, or browser without requiring a public listening port or public IP on the Host.
Key features
Remote workspace access — Open local workspaces from a web browser, …
What problem it solves
Developer workflows often need consistent access to external tools, reusable agent instructions, and service-specific procedures. This repository packages those capabilities as standalone plugins with standardized manifests and supporting rules, skills, and MCP server definitions.
Key features
Developer workflow assistance — Plugins support tasks such as code review, reposit…