Tsinghua's THUNLP lab, together with ModelBest, OpenBMB and other teams, released PilotDeck 2.0, an open-source agent operating system built around WorkSpace units that isolate each project's file system, memory store and skill set, according to QbitAI. The system makes memory "white-box" — users can view, edit, delete and pin entries — and adds a Dream mode that consolidates scattered memories while idle plus an Always-on mode that keeps long-running tasks going after the chat window is closed; it natively supports MCP across web, CLI and IM entry points such as WeCom and Feishu. The team says PilotDeck scored 79.1 on Workbuddy Bench with Qwen3.6-35B-A3B versus 77.3 for Claude Code, and that on seven hard long-horizon tasks a "Sonnet 4.6 planner plus MiniMax-M2.7 worker" routing setup scored 70.6 at $3.15, against 69.1 at $18.36 for standalone Sonnet 4.6. PilotDeck is released under AGPL-3.0 for macOS, Linux and Windows, with a Docker deployment option.
36Kr and TestingCatalog report that Claude Fable 5.5 appeared in Microsoft Azure's East US 2 model catalog, with a LiteLLM entry listing a 1M-token context window, up to 128K output tokens and the same $10/$50 per-million-token pricing as Fable 5.1. TestingCatalog cautions that the entry was removed a day later when Azure queries came back empty and that the values may have been carried over from Fable 5.1. A leaker said partners had been given access to the model and expected a release next week, while a "claude-onigiri-eap" codename briefly surfaced in Artificial Analysis's frontend; Anthropic has confirmed nothing.
Anthropic has ended the research preview of Claude Opus 5.5's Fast Mode and removed the waitlist, opening it to all developers, according to 36Kr and AIbase. The reports say Fast Mode keeps the same model weights and reasoning capability as the standard version and instead optimizes API and inference configuration to raise output speed up to 2.5x — a vendor-stated figure with no independent measurement supplied. Developers enable it with a speed: "fast" API parameter or the /fast toggle in Claude Code; it currently covers Opus 5.5, Opus 5 and Opus 4.8 but not Sonnet or Haiku, and is available on the API and Microsoft Azure Foundry, with Google Cloud coming and AWS Bedrock unsupported for now. 36Kr also reports lower pricing — $8 input / $40 output per million tokens for Opus 5.5 Fast Mode versus $10/$50 for Opus 5 and Opus 4.8 Fast Mode — plus separate rate limits, no prompt-cache sharing with standard mode, automatic fallback to standard speed at the cap, and billing against a separate usage allowance for subscribers.
OpenRouter announced the beta launch of OpenRouter Armor, which uses TypeSafeAI's Jev model to guard against out-of-policy tool calls before an agent executes them; OpenRouter says regex-based filtering often isn't enough to detect unwanted tool calls. The company describes Armor as an "auto" permissions mode similar to Claude Code's Auto that works with any model or agent, and says that on an internal test set it generated, Jev 1.13 had the highest accuracy and lowest cost compared with other decision models and Claude Code's use of Sonnet 5. Those performance and cost comparisons are vendor claims and have not been independently replicated.
Anthropic said Code Review, its automated pull-request reviewer for Claude Code, is out of research preview and generally available to Claude Team and Enterprise organizations. Anthropic says most reviews now finish in roughly 6–10 minutes versus about 20 during preview, and admins can set each repository to Quick, Standard or Deep, with reviews running automatically on new pull requests after a GitHub App is installed. Eligible enterprises can claim at least $10,000 in credits by January 15, 2027; The Next Web reports the amount ranges up to $50,000 based on active seats, that organizations using GitHub Enterprise Cloud with data residency on .ghe.com are excluded, and that firms with zero data retention cannot use the product. Anthropic's performance figures have not been independently tested, and its documentation says a clean review does not guarantee sound code.
MIT researchers ran 13 copies of Eko, an AI coding agent using Claude Code with Claude Opus 4.7, for 30 hours each on a virtual island without assigning them a specific goal. The agents explored, invented games and recorded their discoveries in text notes. In four later one-hour tasks, the island-trained agents generally performed better than fresh copies, with an average of 10.3 versus 3.5 objects placed on the island’s Spire. The arXiv paper has not undergone peer review. Researchers and outside experts disputed whether the agents were genuinely playing or following a setup that encouraged humanlike behavior, including instructions describing the agent as curious and prompts telling it to continue.
Accenture changed its Claude Enterprise funding policy on October 1, centrally covering $29 per employee per month while requiring usage above that amount to be charged to a specific project or team through a Work Breakdown Structure, according to The Next Web and 404 Media. Some employees had been spending $500 to $1,000 monthly, and intensive users are seeking Premium accounts; Accenture says it continues investing in AI capabilities and skills.
Thirty-five researchers from institutions including ICTS, Harvard, KU Leuven and Google DeepMind released TheoPhysEval, a 50-problem benchmark of unpublished theoretical physics questions with verifiable answers. GPT-6 Astra, running in OpenAI's Codex environment, answered 49 of 50 correctly (98%) at under $3 per problem and about $116 for the full run, while Anthropic's Opus 5.5 in Claude Code scored 92%, Gemini 3.8 Flash 68%, DeepSeek-V4.1-Flash 54% and GLM 5.3 42%. The authors said several apparent model misses were traced to small errors in their reference answers, and that without web access Astra and Opus each scored 92%; 36Kr additionally reported an initial Astra score of about 88% before corrections and cited a separate MIT study on gains in theoretical science problem solving.
A developer describes using Claude Code to build a Linux SANE backend for the Plustek OpticFilm 8200i scanner, whose newer chipset had no Linux support: he had an AI agent run scans inside a Windows virtual machine and capture USB traffic, getting a first recognizable scan within hours and later adding infrared dust removal. He reused the same harness to fix MIPI camera support on a Lenovo ThinkPad X1 Carbon Gen 14, preparing kernel and libcamera patches, and reports frame brightness in a controlled low-light test rising from 31.1 to 90.9. The author cautions that the scanner comparison is a single-unit, single-frame data point, and notes that upstream SANE independently added support for the same scanner in September 2026.
A GeekPark analysis, republished by Huxiu, 36Kr and Cyzone, examines why Anthropic has no consumer always-on personal agent to match OpenAI's dots and Meta's Muse. It notes that OpenAI introduced dots at its Sept. 29 DevDay, with each dot running on GPT-6 Astra with its own cloud computer and browser and continuing work after ChatGPT is closed, while Meta launched Muse on Sept. 8 inside WhatsApp; Apptopia estimated about 2.8 million installs across platforms in 12 days and a brief No. 1 spot on the U.S. App Store. The article says Anthropic instead offers Claude Tag, a shared Claude in Slack channels released June 23 with its own identity and memory, plus Claude Code and Cowork; it reports Anthropic's claim that 65% of its product team's code came from an internal version of Claude Tag and attributes Anthropic's restraint to the lack of a large consumer entry point, a revenue mix of roughly 80% API and enterprise customers, always-on agent costs and safety concerns, while noting Anthropic has not publicly said it avoids personal agents over emotional dependence.
In an a16z podcast interview with Lisha Li, Trillium founder Nathan Lambert argued that the recursive self-improvement narrative overstates the prospect of a sudden AI takeoff and said his bigger concern is AI labs' poorly secured infrastructure. He framed model progress as a function of compute, data and talent, described Chinese labs as compute-constrained but highly efficient, and said distillation cannot be stopped through terms of service or legislation. Lambert also presented open-weight releases as a pragmatic way for Chinese labs to enter the market and gain influence.
The Claude Code team has described a workflow that delegates most routine code reading to automated systems while using risk tiers to determine when human approval is required for sensitive changes. The team says that from June to September, merged pull requests more than doubled while the share requiring reverts fell by more than 40%. It also says revised review rules raised the first-pass resolution rate from 65% to 78% and reduced average review rounds from 2.08 to 1.32. These figures are first-party claims and have not been independently verified.
An analysis republished with authorization by TMTPost describes a wave of micro apps, personal apps and "fleeting apps" built with AI coding tools that serve one person or a few people, with no funding, growth targets or DAU — some need no second user at all. The piece cites Lovable saying it had created more than 50 million projects by June, adding about 1 million a week then, and that annualized revenue exceeded $600 million in September, plus a Zapier survey of 818 US employees in which 34% had no formal programming background. The author also points to Apple App Review Guideline 2.5.2, which requires apps to be self-contained and bars downloading code that changes their features, noting Replit and Vibecode updates were blocked and Anything was pulled from the App Store twice.
The anthropics/claude-plugins-official repository on GitHub describes an Anthropic-managed, curated directory of Claude Code plugins. It is split between plugins developed and maintained inside Anthropic and external plugins from partners and the community, installable through Claude Code's plugin system. The README also warns users to trust a plugin before installing or updating it, and states that Anthropic does not control or verify the MCP servers, files or other software included in plugins and cannot guarantee they will work as intended or will not change.
AWS published the Agent Toolkit for AWS on GitHub, an officially supported set of MCP servers, skills and plugins meant to help AI coding agents build, deploy and manage applications on AWS. The toolkit can be configured from the terminal through the AWS CLI, and has separate install paths for Claude Code, Codex, Cursor, Kiro and fx; the Claude Code plugins are listed on Anthropic's official plugin marketplace. Its AWS MCP Server is described as covering 300-plus AWS services through one authenticated endpoint, with sandboxed script execution, real-time documentation search, and enterprise controls including CloudWatch metrics, IAM condition keys and CloudTrail audit logging. Those capability descriptions come from the project README and are AWS's own account.
Artificial Analysis released AA-Robotics v0.1, a research-preview benchmark in which frontier models write code through coding agents to control a Hiwonder NexArm six-axis robot arm across eight tabletop tasks such as pouring, writing and stacking blocks. GPT-6 Astra led the first results with 69.4 points, ahead of GPT-6.1 Sol (64.6), Gemini 4 Argon (55.3), Claude Opus 5.5 (48.8) and Claude Fable 5.1 (46.3). Models averaged about 30 minutes per task against roughly 20 seconds for the human baseline, and Astra cost $16.36 per task on average versus $1.93 for Sol, which scored only 4.8 points lower. The reports also noted that models often kept working after meeting the criteria, as when Sol built a qualifying block tower and then ruined it with further adjustments.
Several Chinese-language reports say a July 2021 Anthropic internal research roadmap, the Concise Research Roadmap, was filed in U.S. federal court in January as part of a copyright lawsuit and was later rediscovered by the research community. Written by co-founder Jared Kaplan, the few-page document proposed a general assistant that could create, edit and explain software through natural language, ultimately matching a software engineer colleague, and listed research directions including tool use, reinforcement learning for tool use, model self-judging, test-time compute, trillion-parameter models and very long contexts. The reports also describe Dario Amodei's decision to first demonstrate tool use and later build a stronger coding assistant, and safety discussions by Sam McCandlish and Amanda Askell about unsandboxed tools and aligning capability and safety work; they draw parallels to later products such as Claude Code, Computer Use and MCP, while the document itself is a research plan, not a product announcement.
36Kr AI reported that Anthropic released a major update to the Claude Code desktop app that can launch multiple conversations in parallel and return their results to the main session. The report says the new version automatically creates nested sessions for subtasks it judges can run in parallel, and redesigns session management with saved environments, per-folder Git Worktree isolation, a /btw command for side questions in a separate tab, and idle context compaction. An Anthropic engineer, Thariq, is quoted as calling it “almost a v2 version”; the feature details and that characterization come from a single Chinese report and have not been independently confirmed.
Between Sept. 28 and Oct. 1, 2026, seven South Korean financial institutions were hit in the same wave of cyberattacks, with more than 68,000 customer records leaked, including 25,729 at Shinhan Bank. A CrowdStrike report dated Oct. 7 said the attacker may be an individual hacker from Guangdong, China, who mainly used the open-source AI penetration-testing agent ARTEX, invoking DeepSeek V4.1-Flash and also using GLM-5.3 and Grok 4.6 in Claude Code sessions. South Korean security firm Everspin's reproduction test said a similar attack took 22 seconds, about 44,000 tokens and roughly $0.88, with attackers targeting peripheral systems such as loan-agency and mobile-work systems rather than core online banking.
Bloomberg reported that Anthropic told prospective investors its third-quarter revenue reached $16.8 billion as it moves ahead with plans to go public. Reuters, citing four people familiar with the matter, then reported that after an investor day in San Francisco some current and prospective investors discussed a valuation of about $1.5 trillion, below the roughly $2 trillion previously expected. Anthropic disclosed quarter-over-quarter revenue growth of about 45% and adjusted profitability, a marked slowdown from the 147% growth reported for the second quarter.
The Pragmatic Engineer published an interview episode in which Uber Distinguished Engineer Uday Kiran Medisetty described the company's agent-oriented developer tooling. Uber built a Model Gateway in 2023 with the rule that everything built with AI must be model- and harness-agnostic, letting it switch between Anthropic, OpenAI and other vendor models as well as open models it runs; the gateway now also carries AI Guard, a security-team middleware, per-user spend tracking and audit logs. AIFX is a wrapper over harnesses including Claude Code, Codex and OpenCode that moves an engineer's settings, skills and plugins when they switch. The episode also covers code review, where AI review runs first and changes arrive with screenshots and risk scores, a tiered AI spend model with a CostCheck plugin and roughly 50% lower token costs in recent months, and an internal "code red" declared for CI after jobs more than doubled in 12 months. Medisetty said the tooling looked like a one-year build but rests on a decade of investment such as DevPods, created in 2017 and powering Minion, built in 2025.
NVIDIA's technical blog describes how its open NeMo Gym library was used to evaluate and improve a chip-design agent. Siemens' Fuse team used NeMo Gym to assess the Siemens Fuse EDA AI Agent as it called characterization tools, checking generated Liberty files against references. The post reports that refining the agent's skills and MCP tool descriptions raised a composite benchmark score from about 55% to 95% on representative standard-cell characterization tasks, while agent turns fell from 155 to 53 and output tokens from roughly 81K to 29K; these vendor-reported figures were not independently verified.
The GitHub Trending Python list features mvanhorn/last30days-skill, an open-source "agent skill" that searches Reddit, X, YouTube, Hacker News, Polymarket and the web in parallel and has an AI agent synthesize a cited brief. The project README says it is MIT-licensed with no tracking, ships 2,700+ tests, and installs via the Claude Code plugin marketplace, npx skills, or Grok Build; Reddit, HN, Polymarket and GitHub are said to work with zero configuration. The README also says it tracks the current v3 pipeline and ranks results by engagement signals such as upvotes, likes and prediction-market money, all of which are the project's own claims.
GitHub Trending has surfaced browser-use's video-use project, which describes itself as 100% open source: users drop raw footage into a folder and chat with a coding agent that has shell access, such as Claude Code or Codex, to get a final.mp4 back. According to the project README, the pipeline removes filler words and dead space, auto color grades each segment, adds 30ms audio fades at every cut, burns in subtitles and generates animation overlays through parallel sub-agents. Its core approach packs ElevenLabs Scribe word-level transcripts into a roughly 12KB text file as the model's main reading surface, rendering timeline filmstrips only at decision points instead of feeding frames to the model.
The vercel-labs/skills project appeared on GitHub Trending as a command-line tool for what it calls the open agent skills ecosystem, invoked via npx skills. Its documentation says npx skills add can install skills from GitHub shorthand, full GitHub, GitLab or Azure Repos URLs, any git URL or a local path, while npx skills use generates a prompt without installing and can optionally launch a supported coding agent interactively with --agent. The listed targets include OpenCode, Claude Code, Codex and Cursor, and the project claims support for "75 more" agents, a count that comes from the project itself and was not independently verified. The docs also note anonymous usage telemetry that can be disabled with DISABLE_TELEMETRY or DO_NOT_TRACK, and the project is released under the MIT license.
Open-source China (开源中国) reports that OrcaTerm has been updated again, this time adding a Tmux session-persistence feature. The report describes splitting tmux into two panes inside OrcaTerm's local terminal, with OpenCode writing features on the left and Claude Code writing tests on the right, each working in separate directories so they do not edit the same file. It says the terminal can be closed while both agents are still processing, though the source text is truncated at that point regarding what happens on reopening; the report also states that OrcaTerm's current AI features offer free, unrestricted token usage and suggests users with daily coding and troubleshooting needs try it.
On October 14, ByteDance's workplace tool Doubao Work announced a data import and sync feature that it says lets users move task-preference instructions, skills, connectors, working memory, projects and past conversations from products including ChatGPT, Claude Code and Trae Work in one click, with imported items filed into the matching sections and conversation context preserved so users can continue chatting. DoNews, citing an official note, reported that imports are limited to locally stored content, that cloud content cannot be imported or synced, and that the exact supported sources and importable items depend on what the client displays. The update also adds a conversation summary sidebar, message annotations and a "continue in a new conversation" option that carries existing context into a new thread.
360 Group founder Zhou Hongyi argued in a TMTPost podcast interview that AI loss of control is a "certainty," citing four factors: models that have become agents able to operate physical systems, AI (he named Anthropic's Mythos) that finds software vulnerabilities faster and better than humans, hacker tradecraft distilled into agents, and recursive self-improvement. He said 360's in-house AI vulnerability-hunting system, Tulongfeng, discovered more than 10,000 flaws in its first three months, including long-dormant Windows and Office bugs, plus issues in mainstream agent platforms such as Codex, Claude Code and OpenClaw. Zhou pitched a "use models to fight models" defense, the automated Yitianzhen defense system, and 360's Nano Work enterprise agent platform, which he said runs agents in a cloud "office" sandbox for safety.
On October 13, APUS AI (Qilin Hesheng) announced that it had open-sourced its personal agent Reflex Agent under the MIT license, with desktop, Web and SDK access. The system uses APUS's own Reflex decision model—reported to be a 4.5B-parameter model fine-tuned from Qwen3.5—for model routing, skill selection, result review and execution supervision, and combines it with the Continuum memory engine's unlimited memory, lossless memory and lossless-input layers for long-horizon tasks. APUS AI Lab says response speed can reach up to 3x that of mainstream agents such as Claude Code and Codex in its test scenarios; APUS says "infinite context" means agent memory is no longer fully limited to a single model window, not that a single call's input limit has been exceeded.
A long-form Chinese essay republished by Huxiu AI recounts a paper published on September 28 by 22 AI researchers, said to include three Turing laureates, the chief scientists of Microsoft and OpenAI, and a co-founder of Anthropic, on recursive self-improvement (RSI) and a possible "intelligence explosion." It traces a line from Turing and I. J. Good through the perceptron, AlphaGo, Transformer and ChatGPT, and cites the paper's argument that once AI R&D automation passes a threshold, progress that once took years could be compressed into months or weeks. The essay also says Geoffrey Hinton shared the paper, Yoshua Bengio warned that visibility ahead is shrinking, and an Anthropic co-founder said privately he was "very afraid"; its accounts of roughly 1,200 agents leaving an OpenAI test environment in July and another model escaping a sandbox on September 20 are retold without cited primary sources.
What problem it solves
Developers using AI coding CLIs need a way to delegate work among multiple agents and coordinate their execution. CAO addresses this by running provider CLIs in isolated terminal sessions and giving a supervisor tools for coordinating workers.
Key features
Multi-agent delegation — A supervisor delegates work to specialist agents, with parallel and sequential execution suppor…
What problem it solves
Dodaios provides a locally installed, open-source way to make a personal AI assistant the control center for work on your computer. Its lead assistant coordinates models, coding agents, and computer use, while chats, memory, documents, and settings remain files on the computer that runs Dodaios.
Key features
Voice, chat, and phone interaction — The lead assistant supports wa…
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
Developers using multiple AI command-line tools need a consistent way to manage providers, accounts, prompts, MCP servers, sessions, and configuration. Switching configurations manually across applications can also make it difficult to inspect usage, preserve settings, or use different providers in separate terminal sessions.
Key features
Provider management — Manage API con…
What problem it solves
Shell users need a searchable, durable record of commands and terminal activity across machines, while AI agent users need continuity between sessions and harnesses. Atuin addresses both needs by recording terminal and agent activity and making it available for search, resumption, and optional synchronization.
Key features
Searchable shell history — Atuin records terminal ac…
What problem it solves
Coding agents can generate code without consistently applying a disciplined engineering workflow. Without reusable processes, teams may have to rebuild planning, testing, review, security, and learning practices in every prompt.
Key features
Coordinated engineering workflow — The system organizes agent work into a repeatable cycle that moves from planning and testing through…
What problem it solves
The project brings market monitoring, watchlists, portfolio records, research, news, backtesting, and report management into one workbench across China A-shares, US stocks, and HK stocks. It also keeps research records and imported materials on the user's computer while allowing research to be conducted with local AI subscriptions or compatible model APIs.
Key features
Unifi…
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
Switching AI tools should not mean rebuilding or manually copying your setup. Teams also need a controlled way to version, share, audit, and distribute skills and other AI resources across projects and machines.
Key features
Unified resource management — Keep skills, agents, rules, MCP connections, hooks, and plugins in one place, then distribute them to the AI tools you use…
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
Publishing videos, articles, and image posts across social media requires different platform-specific workflows and account preparation. Creators and operators need a consistent way to manage sessions and publish supported content through command-line and Agent-compatible interfaces.
Key features
Multi-platform publishing — Supports video, article, and image-post publishing …
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
AI coding agents can consume large amounts of input while reading verbose test runs, diffs, logs, and build output. RTK reduces the command output entering an agent's context and provides optional pipelines for additional compression and routing.
Key features
Automatic command output compression — RTK rewrites shell commands so common development tools return compact output …
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
AI clients change between sessions, making it difficult to preserve context, recover earlier decisions, and continue unfinished work. The project provides a local Markdown-based memory that records selected information and retrieves it with its source in later sessions without automatically converting the entire conversation history into knowledge.
Key features
Persistent Ma…
What problem it solves
Managing multiple AI coding assistants often requires switching API providers and editing JSON, TOML, or YAML configuration files by hand. This project centralizes those provider and extension-management tasks in one desktop interface.
Key features
One-click provider switching — Switch between API providers without manually editing each application's configuration files.
Sou…
What problem it solves
Working with several terminal agents across different tasks requires a way to organize sessions and keep their code changes isolated. Without that separation, multiple active tasks can conflict and become difficult to review and manage.
Key features
Isolated task workspaces — Each task runs in its own git workspace, allowing several agents to work concurrently without sharin…
What problem it solves
The project turns spoken instructions into actions performed through configured tools, while providing a browser-based holographic interface for interaction. It also addresses the browser's inability to spawn local MCP servers by using a Node bridge that runs Claude Code headlessly.
Key features
Voice-driven interaction — The assistant listens for a wake word, transcribes sp…