Claude Design is here! The barrier for designers has dropped to 0.

Claude has come to “kill” another industry.

Yesterday, AI unicorn Anthropic dropped another bombshell by launching its new product, Claude Design, allowing users to create designs, interactive prototypes, PowerPoint slides, single-page documents, and other visual works directly using Claude.

Claude generates an initial draft based on your requirements. Users can then refine it through conversation, inline comments, direct editing, and custom sliders.

The results generated by Claude Design can be exported anywhere — imported into Canva, saved as PPT, PPTX, HTML, or other formats — and can even be packaged with one click into Claude Code to generate the corresponding code.

In fact, Anthropic had already experimented in the design space earlier. Claude Code once released a Skill called “frontend-design” specifically to optimize the web design output of Claude Code.

Now, backed by the much more visually powerful Opus 4.7 model, the Anthropic Labs team has naturally released the full product: Claude Design.

The capital market’s reaction was as brutal as always. Right after the announcement, the stock price of Figma, the U.S. design software leader, dropped 6.84%.

Read more

Just now, OpenAI completely blew up Codex!

Yesterday, you were still using Codex to write code.
Today, it can already see your screen, control your mouse, remember your preferences from last week, and schedule tasks for itself.

Multiple AI Agents are now working in the background for you, without interfering with your mouse or keyboard at all.

Starting today, the tool that 3 million developers use every week is no longer just a coding Agent.

You do your thing.

Now Codex has its own independent cursor that doesn’t interfere with yours.
While you’re writing documentation, you can be running Xcode and testing your App right beside you — both happening at the same time.

What exactly can it do? One demo says it all.

Read more

Unexpectedly, Claude, which has been at the center of attention, has received another major update.

On Thursday, Anthropic announced the full release of its latest flagship model: Claude Opus 4.7.

Opus 4.7 shows significant improvements over Opus 4.6 in advanced software engineering, particularly when handling the most complex tasks. Users report that they can now confidently hand over even the most challenging coding work—tasks that previously required close supervision—to Opus 4.7. The model handles complex, time-consuming tasks with rigor and consistency, executes instructions with high precision, and designs methods to verify its own outputs before returning results.

The model also delivers noticeably better visual performance: it can process higher-resolution images and demonstrates greater taste and creativity when performing professional tasks, producing higher-quality interfaces, slides, and documents. Although its capabilities are not as comprehensive as the newly announced strongest model, Claude Mythos Preview, Opus 4.7 outperforms Opus 4.6 on multiple benchmarks:

Read more

Anthropic Official In-Depth Guide: Multi-Agent Collaboration Patterns

This guide summarizes 5 architectures along with their ideal use cases, when you should use a multi-agent system, and when a single-agent approach is actually the better choice. If you’ve already decided to go with multi-agent systems, this article is exactly what you need.

We often see development teams blindly adopt certain architectures simply because they sound advanced or impressive — which is a big mistake. Anthropic strongly recommends starting with the simplest possible mode that can actually get the job done, carefully observing where it hits bottlenecks, and then evolving the system step by step. Today, we’ll thoroughly break down the underlying mechanisms and critical limitations of these five multi-agent collaboration patterns.

The five patterns are:

1. Generator + Verifier – For output quality control with clear, well-defined standards

2. Orchestrator + Sub-Agents – For clearly breaking down independent subtasks

3. Agent Team – For parallel, long-running, independent work

4. Message Bus – For event-driven pipelines and expanding ecosystems

5. Shared State – For collaborative research and knowledge sharing

Read more

Anthropic has officially launched Routines, the automated task feature for Claude Code. It is currently in research preview.

Once you configure the prompt, code repository, and connectors just once, Claude can work completely autonomously in the cloud. All tasks run on Anthropic’s cloud infrastructure, meaning you don’t need to keep your computer on at all. Even if you shut down your machine after work, it can still handle your code backlog, review code, and respond to cloud events on schedule.

This feature completely eliminates the previous cumbersome process where developers had to manually manage cron jobs, set up infrastructure, and maintain MCP servers. The system supports three highly flexible triggering methods that cover almost all daily development scenarios.

Read more

A dedicated Claude Code configuration repository on GitHub has gone viral, racking up 154K stars. The author spent 10 months refining it and won the Anthropic Hacker Marathon Grand Prize. The repository is called Everything Claude Code.

This repo contains a total of 47 sub-agents, 181 skills, and 79 commands. It is far more than just a set of configuration files — it’s a complete overhaul of how you work with AI.

The author, Affaan, said these configurations were battle-tested over more than ten months in real product development. After winning Anthropic’s Hacker Marathon, he open-sourced everything.

I spent an afternoon dissecting the entire repository. Here’s my understanding:

Read more

Hermes may be a better fit for you than OpenClaw.

If you care more about the agent’s long-term evolutionary capabilities and want it to become smarter the more you use it, or if you are doing AI research and need to generate training trajectories or run reinforcement learning experiments, Hermes’ architecture is a much better match.

Hermes has a built-in server compatible with the OpenAI API, so it can be directly used as a backend for third-party interfaces such as Open WebUI.

The deployment threshold is also very low. A $5/month VPS is sufficient to run it. It supports Docker, SSH remote access, and serverless solutions like Modal. Installation can be completed with a single curl command.

Below is the installation process for Hermes Agent and my actual usage experience over the past few days.

Taking the Mac platform as an example.

Installing Hermes Agent only requires one command:

curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.sh | bash

Read more

Anthropic Has Made Another Move: Implementing a “Advisor Strategy” on the Claude Platform

The main goals are to reduce costs and boost the intelligence of lighter models. Recently, many users have reported that Claude seems to have become “dumber.” It’s increasingly clear that Anthropic’s top-tier models are prioritizing B2B and internal iteration use cases, becoming less friendly to individual C-end (consumer) users. One has to wonder whether Anthropic has been secretly using this kind of “Advisor Strategy” internally for a while already.

In simple terms, the most powerful Opus model acts as an Advisor in the background, while a lightweight Sonnet or Haiku model serves as the Executor. With this combination, developers can instantly give their agents near-Opus-level intelligence at extremely low cost.

Read more

Anthropic has released a new product called Claude Managed Agents.

My first reaction upon seeing the announcement wasn’t excitement — it was relief.

Anyone who has built Agents knows that the real pain has never been whether the model is smart enough.
The models are already good enough. What actually hurts is building the entire infrastructure layer around the model: sandbox environments, credential management, task recovery from failures, permission isolation, and end-to-end tracing.

None of these things is particularly difficult on its own, but when you have to put them all together, run them reliably in production, and handle real user traffic, it suddenly becomes several months of engineering work — not just a few weeks.

This leaves you with two choices: either spend those months building it properly, or cut corners and ship something that works as a demo but probably shouldn’t go to production.

Read more

Every month, GitHub gives you a bunch of Copilot quota, but most people use up only a tiny fraction of it.

As a subscriber paying $10 or more, you also get nearly unlimited free access to GPT-4o and other models.

Since you probably can’t exhaust the quota just by using Copilot in your editor, you can proxy this service and share it with other Chat tools or OpenClaw.

I’m currently using Coxy, a local proxy service that translates GitHub Copilot’s request protocol into a standard OpenAI-compatible API.

Once started, it exposes a local endpoint at http://localhost:3000/api. Any tool that supports the OpenAI API can connect to it, including Open WebUI, LLM CLI, or your own scripts.

Read more

Last month, Anthropic’s most powerful model, Claude Mythos, was unexpectedly leaked.

The leaked internal documents stated that it is larger and more intelligent than Anthropic’s Opus model and is the most powerful AI model ever developed.

Afterward, Anthropic attributed the leak to “human error.”

And just now, this “leaked” model has officially launched, along with a much larger plan. In the past, we generally believed that the threat of AI came from it being “too stupid”: hallucinations, errors, and unreliability. Today, Mythos brings a different kind of panic: it is too smart.

AI’s ability to find vulnerabilities has already surpassed that of the vast majority of humans.

Anthropic, together with AWS, Apple, Microsoft, Google, NVIDIA, Cisco, Broadcom, CrowdStrike, JPMorgan Chase, the Linux Foundation, and Palo Alto Networks — a total of 12 organizations — has launched the Project Glasswing initiative.

These 12 organizations cover almost every aspect of global digital infrastructure: operating systems, chips, cloud computing, cybersecurity, financial infrastructure, and open-source ecosystems. Nothing was left out.

Newton Cheng, Anthropic’s Head of Frontier Red Teaming for Cybersecurity, said: “The reason we created Glasswing is to give defenders the first-mover advantage.”

Read more

Anthropic’s annualized revenue has surged by approximately $21 billion in just over three months.

Today, Anthropic announced in a blog post that its annualized revenue (ARR) — calculated based on recent actual revenue run rate — has exceeded $30 billion, surpassing OpenAI, whose latest annualized revenue stands at $25 billion.

At the end of 2025, Anthropic’s annualized revenue was approximately $9 billion. This means the company has achieved an explosive 233% growth in just over three months.

Read more

Codex CLI is a coding agent tool launched by OpenAI, with over 90% of its core code written in Rust.

This article compiles best practices for the full Codex workflow — from configuration to orchestration — and is written for developers who want to deeply integrate Codex CLI into their daily engineering process.

This is a no-fluff version of everyday practices, covering content from beginner to advanced, so you can truly master the most important parts of Codex, pick it up immediately, and use it well from the start.

Core Concepts

Before diving into specifics, let’s establish an overall map.

The capabilities of Codex CLI can be broken down into the following layers:

Command Layer

Built-in commands are triggered by a slash in an interactive session.

These include: /plan, /fast, /fork, /review, /status, /mcp, /agent, /model, /permissions, and others — used to control the behavior of the current session.

Agent Layer

Uses TOML files in the .codex/agents/ directory as configuration units, defining customized agents with dedicated roles, models, and instructions. Supports parallel sub-agent orchestration and CSV batch processing.

Skill Layer

Reusable instruction packages in the standard format .agents/skills/<name>/SKILL.md, invoked explicitly by name or triggered automatically via description fields.

Plugin Layer

Skills and MCP servers packaged with application integrations into distributable units, installable via a plugin marketplace or local path.

Configuration Layer

A TOML-based layered configuration system supporting sandbox modes, approval policies, model switching, and MCP server registration.

Hook Layer

Injects custom scripts into various nodes of the agent loop, enabling deterministic automation such as logging, security scanning, and format validation.

All major workflows converge on the same architectural pattern: Research → Plan → Execute → Review → Deliver.

Understanding this central axis is the prerequisite for using Codex CLI effectively.

Read more

Over the past two days, more and more information about GPT-6 has been leaking out, and Sam Altman’s anxiety has reached its peak.

Although OpenAI hasn’t released any major products since entering 2026, Altman’s recent schedule has been explosive:

After killing off the flagship product Sora, apologizing to Disney’s CEO over the phone, and finishing a heated argument with the CFO about IPO timing, he had to put on the suit he’s worn more than any other in his life to meet with the head of state of a certain country to discuss chips. Upon returning, he discovered that $600 million worth of OpenAI secondary market shares had been listed for several weeks with no buyers.

Amid internal troubles and external pressures, GPT-6 is a trump card that OpenAI simply cannot afford to lose.

This internally codenamed “Spud” (potato) model has undergone two years of secret development. Pre-training was completed on March 24 at the Stargate data center in Texas, with the earliest expected release in April.

OpenAI President Greg Brockman recently confirmed its existence in a podcast interview: “This is not an incremental improvement, but a major change in how we think about model development.”

On the same day, Altman raised expectations even higher in his address to all employees: “A very powerful model that can truly accelerate economic development.”

OpenAI has been holding back this major move for two years, and now it’s about to release it—at the most dangerous moment in the company’s history.

Read more

OpenClaw Founder: Begging Only Bought 7 Days of Extra Life

Anthropic announced that starting at 3:00 PM Eastern Time on April 4, Claude has disabled third-party tools such as OpenClaw. Users can now only access these tools through additional paid packages or by using a Claude API Key.

The official email sent by Claude also stated that this policy will first be strictly enforced on OpenClaw starting April 4, but it applies to all third-party toolchains and will soon be extended to more tools.

The core reason for the blockade is that Anthropic wants to prioritize and protect customers who use its own products and API.

Read more

The most powerful CFO in the tech industry personally slammed the brakes on data center expansion last year. Now, the compute shortage she helped create is biting back at Microsoft’s growth, while the projects she abandoned have been snapped up by competitors. Is this woman—who wields godfather-level influence on Wall Street—playing a grand strategic game, or has she made a costly misjudgment?

According to Bloomberg, Microsoft CFO Amy Hood raised serious doubts about the company’s internal capital expenditure forecasts at the end of 2024. At the time, Microsoft’s quarterly spending had already surpassed the total for the entire previous year. The company was building data centers, stockpiling chips, and preparing for an explosion in demand for AI and cloud services at an unprecedented pace.

Read more

This Thursday, Google open-sourced the Gemma 4 series, currently the strongest model family in the open-source world.

Built on the same research breakthroughs as Gemini 3, the new models achieved 3rd place globally on the Arena AI leaderboard, surpassing models with 20 times more parameters. More importantly, this generation of Gemma uses the Apache 2.0 open-source license, enabling complete commercial freedom.

Gemma 4 is Google DeepMind’s latest series of open models. They are multimodal models capable of processing text and image inputs (with small models also supporting audio input) and generating text output. This release includes both pre-trained and instruction-tuned open-weight models. Gemma 4 features a context window of up to 256,000 tokens and supports over 140 languages.

Google stated that Gemma 4 employs both dense and Mixture-of-Experts (MoE) architectures, making it highly suitable for tasks such as text generation, coding, and reasoning. The series includes four different sizes: E2B, E4B, 26B, A4B, and 31B. Thanks to their varying sizes, these models can be deployed across a wide range of environments—from high-end smartphones and laptops to servers—allowing more people to access cutting-edge AI.

Notably, the largest 31B version can perform full-precision inference on a single 80GB H100 GPU, demonstrating performance comparable to Qwen 3.5 397B.

Read more

The source code of Claude Code has been leaked. It contains 1,884 files and is 33MB in size.

The leak was most likely caused by a bug in Bun.

Some people also claim it was done intentionally by Anthropic right before a major version update. This is because Claude Code had already been leaked once before, and reverse-engineering alone didn’t yield much useful information.

Regardless of how it happened, the actual source code of Claude Code is now sitting right in front of us.

Beyond the previous summary of what we saw after the Claude Code source code leak—such as unreleased new features and the full future roadmap—the biggest significance of this leak is that it allows us to truly understand Claude Code for the first time.

Before this, all our understanding and usage of Claude Code was based solely on Anthropic’s sparse official documentation and our own speculation.

Now, with the actual source code as reference, the inner workings and operating mechanisms of Claude Code are completely exposed. After reading it, I realized there are several places in my own coding approach that definitely need to be changed.

Read more

Yesterday, the biggest bombshell in the AI world was the “accidental” open-sourcing of Claude Code’s source code.

Due to an engineering mistake, Anthropic failed to strip source map files when publishing an npm package, allowing the complete TypeScript source code to be easily reconstructed. Within just a few hours, the code had been downloaded, mirrored, and rapidly spread across GitHub.

In a way, isn’t this kind of leak a form of open-sourcing? Even Elon Musk, upon seeing someone comment that “Anthropic is now more Open than OpenAI,” couldn’t help but “compliment” it: “Absolutely 😂”

As for the cause, it wasn’t particularly complicated. Although Anthropic has yet to release an official report, tech outlet Decrypt obtained a comment from an Anthropic spokesperson: “Earlier today, a version of Claude Code contained some internal source code. No sensitive customer data or credentials were involved or exposed. This was a human error in the release packaging process and did not constitute a security vulnerability. We are taking steps to prevent this from happening again.”

Claude Code’s creator, Boris Cherny, also briefly stated on X that it was simply “developer error.”

While most people were still watching this dramatic accidental open-sourcing unfold, another group had already started quietly reading through the code line by line, attempting to reconstruct the design logic behind it — and some previously undisclosed system-level strategies were revealed, particularly around model capability protection and data security, where Claude Code clearly involved deeper engineering design.

Read more

I’ve been using Claude Code for a while now. There’s no doubt it’s incredibly powerful, but there’s one thing that has always annoyed me: the way it reads files is far too crude and aggressive.

When you ask it to modify a function, it first reads the entire file. Then it discovers that a dependency is in another file, so it reads that one too. After reading it, it realizes it needs to see the definition of a certain class, so it reads a third file.

For a simple request, it ends up reading three or four entire files. Tokens just pour out. On top of that, with OpenAI cutting three products in a week and Claude once again tightening its usage limits, it feels like the good days of AI are coming to an end. Even the $20 official subscription is often used up after just a couple of sessions.

Actually, this isn’t really a problem with Claude Code specifically—it’s a common disease shared by almost all AI coding agents.

Agents have no concept of a project directory structure. Their default action is: Not sure? Read the file. Still not sure? Read another one.

Recently, I discovered a tool called cx that’s specifically designed to solve this problem. After using it, the results have been excellent, so I’m sharing it with everyone.

Read more