AI Agent News
Track major events, funding, model releases and breakthroughs across the AI Agent landscape
Latest industry news
Track major events, funding, model releases and breakthroughs across the AI Agent landscape
Key events timeline
OpenClaw erupts on GitHub
OpenClaw hits global GitHub Top 10 in 10 days, outpacing the Linux kernel star growth
Meta acquires Manus for $2B
Meta acquires Manus AI for $2B, locking in the general-purpose Agent race
DeepSeek-V3 open-sourced
The value king, at just 5% of GPT-4 cost
Manus goes viral overnight
The world's first general-purpose AI Agent draws unprecedented attention
OpenAI Deep Research
OpenAI ships a deep-research Agent that generates professional reports in one click
MCP Servers pass 500
The MCP ecosystem erupts — 500+ servers built in 3 months
DeepSeek-R1 stuns the world
Open-source reasoning model at just 3% of OpenAI cost, reshaping the global AI landscape
MCP protocol born
Anthropic releases the Model Context Protocol, the de facto standard for Agent interfaces
Claude Computer Use
Anthropic lets AI directly control the computer screen for the first time, opening a new paradigm
Replit Agent full-stack automation
Natural language to a shipped product, aimed at non-engineers
Cursor ARR passes $100M
The fastest-growing SaaS ever, the new king of AI coding tools
Claude 3.5 tops SWE-bench
The strongest coding AI, bug-fixing at a junior engineer level
Devin launches
The world's first autonomous AI software engineer, able to complete full coding tasks on its own
Claude Code Engineer Shares Fable 5 Methodology: Bridging the Information Gap Between Humans and Models
Thariq Shihipar, a core engineer at Claude Code, recently published a detailed blog post outlining the core methodology for using the Fable 5 model. He points out that when model capabilities are strong enough, the bottleneck for work quality shifts from the model itself to the user's ability to clearly articulate requirements. Thariq categorizes unknowns in tasks into four types: known knowns, known unknowns, unknown knowns, and unknown unknowns, emphasizing that 'unknown unknowns' are the most challenging. He proposes a closed-loop process: before implementation, reduce unknowns through blind spot scanning, brainstorming, questioning, providing references, and implementation plans; during implementation, maintain an `implementation-notes.md` to record deviations; after implementation, package the pitch and conduct tests. In a separate interview, Anthropic's Fiona Fung added that the team's per capita quarterly code output has increased 8-fold, but the rise in parallel agents brings new issues such as switching overhead and loneliness.
Anthropic Launches Claude Science: An AI Workbench for Researchers, Claiming 10x Speedup in Tests
On June 30, 2026, Anthropic released Claude Science, an AI workbench designed for scientists—not a new model. It integrates literature search, data analysis, chart generation, manuscript writing, and other research workflows into a single environment, supporting macOS and Linux. Pro, Max, Team, and Enterprise users can access it immediately. ## Core Features and Workflow - **Reproducibility**: Each chart comes with the generating code, runtime environment, and conversation history. Supports natural language modifications (e.g., "remove gridlines"), and results are publication-ready. - **Intelligent Compute Scheduling**: Automatically manages local, SSH remote, or HPC cluster resources, scaling from a single GPU to hundreds of GPUs. Sensitive data never leaves the original system. - **Out-of-the-Box**: Pre-installed with 60+ scientific skills and connectors, directly linking to hundreds of databases such as UniProt, PDB, Ensembl, ClinVar, ChEMBL, and GEO. Integrates NVIDIA BioNeMo Agent Toolkit, supporting specialized models like Evo 2, Boltz-2, and OpenFold3. - **Review Mechanism**: Built-in review agent dedicated to checking citation accuracy and computational correctness. All critical operations require user-by-user authorization, ensuring human-in-the-loop. ## Test Performance - In a brain-computer interface literature review task, Claude Science first formulated a six-step plan, then automatically fetched metadata from OpenAlex after confirmation. When rate-limited, it switched to Crossref. Ultimately, it filtered 1,218 relevant papers (62% with abstracts), performed clustering and trend analysis, and output a Chinese report, charts, and CSV files. - The entire process is visible and traceable, proactively noting data source limitations and suggesting supplementary authoritative databases. ## Early User Feedback - **Allen Institute**: Neuroscientist Jérôme Lecoq's team compressed long-form review writing from nearly 2 years to a few weeks, using an "actor-critic" dual-agent model (one writes, one critiques), producing multiple hundred-page results. - **UCSF Brain Tumor Center**: Glioma molecular epidemiology analysis time reduced to 1/10 of the original, with independently verified reliable results. - **Manifold Bio**: Used for tissue-targeted drug target screening, evaluating surface expression, transport, and safety, outperforming general coding assistants in efficiency. ## Industry Comparison and Cautious Voices - **OpenAI**: Launched GPT-Rosalind (proprietary fine-tuned model, enterprise-only) and GeneBench-Pro benchmark to test model scientific judgment. - **Google**: Holds exclusive models like AlphaFold and AlphaGenome; Gemini for Science integrates 30+ scientific databases. - **Anthropic**: Adopts a subscription-based, open, and workflow-first route, while launching a drug discovery initiative funding 50 AI for Science projects (up to $30,000 each). - **Northeastern University's Jared Auclair** cautions: AI is still a "co-pilot"; it may hallucinate or miss details when interpreting regulatory guidelines or designing experiments, so careful use is advised.
Alibaba Bans Claude Code Internally After Covert Detection of Chinese Users Exposed
On July 3, 2025, Alibaba issued an internal notice banning all Anthropic products, including Claude Sonnet, Opus, Fable models and Claude Code, requiring employees to uninstall by July 10. The alternative is Alibaba's self-developed Qoder. ## Background - Earlier this year, Alibaba encouraged employees to use external AI tools with generous reimbursement, covering Claude, GPT, Gemini, etc. - On June 10, Anthropic wrote to the U.S. Senate Banking Committee, accusing Alibaba of using about 25,000 fake accounts to interact with Claude over 28 million times between April 22 and June 5, labeling it an "industrial-scale model distillation attack" and elevating it to national security. - In late June, many Chinese users had their Anthropic accounts banned, including personal subscriptions and team accounts, with paid accounts not refunded and appeals difficult. ## Key Detail: Claude Code's Covert Detection Mechanism - On June 30, Reddit user LegitMichel777 reverse-engineered Claude Code and found that since version 2.1.91 released on April 2, it had a built-in covert user detection mechanism. - The mechanism works in three steps: - Reads system timezone (e.g., Asia/Shanghai or Asia/Urumqi), checks proxy addresses or custom APIs for 147 Chinese cloud and AI company domains (e.g., Alibaba, ByteDance, Baidu, Moonshot AI, MiniMax). - If triggered, no popup appears; instead, it modifies the system prompt's date format (e.g., changes "2026-06-30" to "2026/06/30") and replaces the apostrophe in "Today's date is" with different Unicode characters (right single quotation mark, modifier letter apostrophe, modifier letter turned comma) corresponding to different hit states. - The modified prompt is sent back to Anthropic servers with normal requests, forming an environment fingerprint. - The core code logic is obfuscated, with 147 domains encrypted and no mention in version changelogs. ## Reactions and Impact - On July 2, Claude Code team member Thariq Shihipar acknowledged the measure as an "experimental" feature launched in March to prevent account resale and model distillation, and claimed it was rolled back in the same day's new version. - On July 3, Alibaba responded the next day by issuing the ban, classifying Claude Code as a security risk with implanted backdoors, adding it to the high-risk software list. - This move marks a shift in China's top tech companies' attitude towards relying on external closed-source tools, from "borrowing" to "security first."
Claude Code Exposed with Steganographic Trojan, Codex Log Bug Writes 640TB Annually: Trust Crisis in AI Coding Tools
Recent severe issues have emerged in AI coding tools. Anthropic's Claude Code has been accused of embedding obfuscated code in binaries since April, using steganography to encode user timezone, proxy, and Chinese AI lab domain names into system prompts to detect unauthorized resale and model distillation. The mechanism employs XOR encryption and Unicode character substitution, making it difficult for users to notice. Claude Code lead Thariq responded that it was an experiment, has merged a rollback PR, and expects to remove it tomorrow. Meanwhile, OpenAI's Codex was found to have a serious logging bug: the default TRACE level logging causes SQLite database writes of up to 640TB per year, enough to burn out a consumer-grade SSD in a year. The issue stems from a single line `with_default(Level::TRACE)` configuration that ignores the `RUST_LOG` environment variable, with at least 9 related issues filed. After the fix, about 15% of writes remain (approximately 96TB per year). Additionally, Codex recently experienced abnormal quota consumption: a single message exhausting the entire quota, background tasks stealing tokens, failed tasks repeatedly retrying generating "ghost quotas", and misaligned usage statistics. OpenAI has reset quotas multiple times and rolled back related changes. These two incidents expose systemic flaws in resource budgeting, user trust, and testing processes for AI coding tools. The community criticizes them as "bad software" and questions vendors using user devices as free debugging storage and monitoring tools.
Claude Code Next-Gen Upgrade and Team Role Restructuring: Background Sub-Agents Become Default, Five New Roles Replace Traditional Positions
Anthropic's AI coding tool Claude Code is getting a major upgrade: sub-agents will run in the background by default, allowing users to chat with Claude while multiple agents execute code refactoring, testing, and PR creation in parallel. Boris Cherny, the creator of Claude Code, also proposed a new team division framework for the AI era, replacing traditional roles (engineer, PM, designer, etc.) with five behavior-based roles: Prototyper, Builder, Sweeper, Grower, and Maintainer. ## Next-Gen Upgrade: Background Sub-Agents Run by Default Boris Cherny announced on X that the next version of Claude Code will enable background sub-agents by default. Users no longer need to manually specify; agents will work in parallel in the background while users continue chatting with Claude. This design upgrades Claude Code from a "question-and-answer dialog box" to a "workflow engine capable of orchestrating multiple task lines." Previously, Claude Code already had several capabilities: - **Routines**: Package prompts, code repositories, and connectors into fixed workflows triggered hourly, nightly, or weekly, or via API, GitHub events, or Webhooks, running on Anthropic's managed cloud. - **Dynamic workflows**: Users just mention "use a workflow" or enable ultracode, and Claude generates orchestration scripts to schedule dozens to hundreds of sub-agents in phases, with parallel cross-validation. This upgrade bundles these capabilities as default behavior, requiring no manual toggling. ## Five New Roles: Traditional Job Labels Removed Boris Cherny proposed on X that as engineering, product, design, and data science functions merge, the Claude Code team no longer uses traditional job titles. Instead, they divide roles based on behavior patterns into five categories: - **Prototyper**: Generates many ideas, most of which won't ship; focuses on quantity and disruption. - **Builder**: Quickly turns prototypes into production-grade products or infrastructure; core is execution speed and engineering judgment. - **Sweeper**: Simplifies UI, streamlines code and system architecture, cuts redundant features, optimizes performance. - **Grower**: Takes over mature products and iterates to improve product-market fit (PMF). - **Maintainer**: Guards mature systems, ensuring security, reliability, and efficiency. Cherny emphasized that these roles are not tied to specific positions. Inside Anthropic, designers, engineers, PMs, and data scientists may fall into different roles, and many people span 2-3 roles simultaneously. Roles change dynamically with project phases: - New products (no PMF): Need many Prototypers, Builders, Sweepers. - Growth products (initial PMF): Shift focus to Builders, Sweepers, Growers, with a few Maintainers. - Mature products (strong PMF): Primarily Sweepers, Growers, Maintainers, with some Builders. ## Real-World Data: Engineer Output Triples, Bottleneck Shifts to Decision-Makers Claude Code's effectiveness has been validated both inside and outside Anthropic: - **Inside Anthropic**: Claude Code contributed 4% of public commits on GitHub, with annualized revenue exceeding $2.5 billion. Each engineer's output is triple the headcount; a five-person team does the work of 15-20 people. Anthropic's growth team decided to hire more PMs rather than engineers because the bottleneck has shifted from "writing code" to "deciding what code to write." - **Spotify**: Engineering VP Niklas Gustavsson revealed that in their monorepo of over 20 million lines of code, there are about 4,500 production deployments daily, 73% of pull requests are AI-assisted, and PR frequency increased by over 75%. He personally runs 5-10 Claude sessions simultaneously, each with an independent git worktree, letting multiple agents work in parallel while he only reviews diffs and makes decisions. Spotify also opens this capability to non-engineers: PMs, designers, and even co-CEOs can describe ideas in natural language, and Claude implements end-to-end prototypes in real code. ## Impact and Discussion: Job Boundaries Melt, Golden Age of Generalists Boris Cherny believes AI is compressing the gap between skills, ushering in a golden age for generalists. He hasn't handwritten code in eight months; all code under his name is 100% generated by Claude. He predicts that in three years, the number of people writing code and using agents will be 100 times today's, but the title "software engineer" may disappear, replaced by more flexible roles. The community response has been positive. One data scientist said they often act as a Sweeper while also bringing data science taste to product building, a "2+3 type." Senior engineer Kun Chen pointed out that roles should change with projects; locking oneself into one role limits growth. Some questioned: if AI can write code, why do we need Builders and Sweepers? Cherny responded that Claude already handles these tasks well and will improve, but humans still need to take responsibility for AI outputs. The Sweeper role is particularly underappreciated. As AI-generated code surges, the proportion of AI code going directly into production without human review has risen from about 7% in early 2026 to nearly 38.5% (per Cursor's Developer Habits Report), increasing hidden system issues and highlighting the Sweeper's value in safeguarding quality.
OpenClaw and Cursor Launch Native Mobile Apps on the Same Day, Bringing AI Agents into the Pocket Era
On June 29, 2026, AI Agent tool OpenClaw and AI coding assistant Cursor both released native mobile applications. OpenClaw launched iOS and Android native apps, enabling remote control of locally deployed AI agents for file operations, email cleanup, automated task approval, and more. It adopts a local-first architecture, with a free tier offering 20 messages per day and a paid tier at $20/month. Cursor released an iOS public beta (requires iOS 26.0+), offering cloud-based Agent and remote control modes, supporting voice input, asynchronous execution, PR merging, and adapting to iOS Live Activities and lock screen notifications. Both apps aim to extend AI Agents from desktop to mobile scenarios, allowing users to initiate or monitor tasks anytime, anywhere, marking a shift in AI autonomy from "operator" to "approver" roles.
Meta Releases Brain2Qwerty v2: Non-Invasive Brain-Computer Interface Achieves 61% Decoding Accuracy
On June 29, 2026, Meta released Brain2Qwerty v2, a non-invasive brain-computer interface system that converts brain activity into full sentences in real time without surgery. The system uses magnetoencephalography (MEG) signals and an end-to-end deep learning architecture combining Conformer, aligner, and a large language model (LLM) for three-level decoding from characters to sentences. ## Technical Breakthroughs - **Decoding Performance**: Average word accuracy of 61% across 9 subjects, with the best subject reaching 78%. Over half of the sentences had only one word or fewer errors. This is a significant improvement over traditional non-invasive methods (~8% accuracy). - **Architecture Upgrade**: v2 upgrades from character-level decoding in v1 to word- and semantic-level decoding, outputting complete sentences directly. The model consists of a Brain Encoder (Conformer module) and NeuroLLM (LLM based on Qwen3-4B), aligning neural features with word embeddings via contrastive loss. - **Training Data**: Each subject wore MEG equipment for about 10 hours, collecting approximately 22,000 sentence samples of imagined typing. Research confirms that decoding accuracy increases logarithmically with data volume. ## Comparison with Invasive Approaches Invasive BCIs (e.g., Neuralink) achieve over 90% decoding accuracy but require craniotomy, posing high risks and limiting accessibility. Brain2Qwerty v2 approaches invasive performance non-invasively, laying the foundation for large-scale applications. ## Open Source and Ecosystem Meta fully open-sourced the training code for v1 and v2, and partner BCBL simultaneously released the v1 dataset. Additionally, Meta released the perceptual encoding Tribev2, brain data processing tool NeuralSet, evaluation platform NeuralBench, and invested $5 million in open brain science datasets. ## Challenges and Outlook - **Insufficient Accuracy**: Current models still have word or character errors, making them unsuitable for daily communication. - **Equipment Limitations**: MEG systems are bulky and expensive (millions of dollars), relying on liquid helium cooling and magnetic shielding. Next-generation portable OPM-MEG is still under development. The technology aims to serve patients with communication disorders such as stroke and ALS, enabling communication without surgery. In the long term, it may drive changes in the diagnosis and treatment of neurological diseases.
TRAE Work Launches Design Mode: Bridging Requirements, Design, and Code
TRAE Work recently launched its Design mode, which, together with the existing Work (requirements) and Code modes, forms an end-to-end workflow from requirements to design to code. This mode supports importing Figma files to automatically extract design systems (colors, fonts, components, etc.), and also allows establishing design context through style exploration or built-in brand libraries. Users can modify designs on the canvas via conversation, box selection, or panel parameters, and export them to Figma or jump directly to Code mode for development with one click. Real-world tests show that Design mode effectively maintains brand consistency, reduces context loss when switching between tools, and compresses the multi-round collaboration process among product, design, and front-end teams to within one hour. However, AI remains conservative in creative aspects such as visual impact, requiring manual supplementation. This mode aims to address common pain points in AI design tools—such as "not recognizing design systems," "weak editing capabilities," and "tool silos"—to promote design asset reuse and workflow integration.
Doubao Pro Launched: Office Agent Capabilities Tested and Pricing Controversy
In June 2026, ByteDance's AI assistant Doubao launched a professional edition with three subscription tiers (68, 200, 500 yuan), featuring an 'office task mode' Agent that autonomously decomposes goals, calls tools, executes tasks, and delivers results. Pricing comparisons with domestic competitors (Zhipu GLM Pro 49-469 yuan, Kimi 49-699 yuan, Xiaomi MiMo 39-659 yuan, ByteDance Trae Work 59-1399 yuan) show domestic AI pro versions approaching $20/month, raising cost concerns. Real-world tests by Z Finance demonstrated strong Agent performance in automatic weekly reports, industry PPTs, no-code website building, financial report analysis, and literature reviews, with accurate data and multi-step execution. However, a complex creative coding test (astronomical clock) revealed stability issues: Doubao 2.1 Pro took 26 minutes but had 5 code errors and failed to run. In contrast, end-to-end engineering delivery tests by Datawhale showed Seed 2.1 Pro successfully built a full-stack paper-graph-manager system from a single requirement, demonstrating robust capabilities in requirement understanding, code engineering, and long-chain Agent execution. Overall, Doubao Pro marks a shift from free growth to monetization for a national-level AI app, but whether its Agent capabilities justify the pricing remains to be seen.