Welcome back. Anthropic is on a hot run. They just dropped what's likely their biggest update of the year, with the launch video already reaching 37M views. Claude can now take over your entire Mac. It can open apps, browse the web, fill out spreadsheets, and click through workflows even when you are not at your desk.

Also: How to review PRs like OpenClaw's creator, get GPT-5.4 to build frontends that look well designed, and find out why Jensen Huang says we don't have AGI yet.

Anthropic gives Claude direct control of your Mac: The AI lab recently released a research preview that allows Claude to take over your desktop and complete tasks from start to finish. It can click, type, open apps, navigate browsers, and edit files. Before using screen control, it checks for secure integrations with apps like Slack and Calendar. This new feature works alongside Dispatch, letting users assign tasks from their phones and return later to find the work already completed.

ChatGPT now saves your files across conversations: The ChatGPT Maker just revealed Library, a new built-in storage feature that automatically saves any documents, spreadsheets, or images you upload or create during a chat. You can find these files in a dedicated sidebar tab on the web version, making it easy to pull them into future conversations without having to re-upload anything. Even if you delete a chat, the associated files will remain saved until you manually remove them.

Cursor makes AI agents 1,000x faster at searching code: The coding startup recently launched Instant Grep, a client-side indexing system capable of searching millions of files in just 13 milliseconds. It works by creating a local index tied to Git commits, utilizing n-grams and Bloom filters to narrow down matches before scanning files directly. This represents a significant improvement over ripgrep, which takes nearly 17 seconds for the same query. These speed gains are especially helpful in large monorepos, where slow searches have been a major bottleneck for AI coding agents.

The bottleneck isn't the model. It's typing detailed context 50x a day.

Wispr Flow lets you speak full prompts into Cursor, Windsurf, or ChatGPT with syntax-aware dictation that respects camelCase, snake_case, and file names.

Fewer reprompts. More shipping.

Try Wispr Flow Free

The line is getting clearer. When Andrej Karpathy coined the term "vibe coding" in February 2025, it captured something real: prompting an LLM and simply accepting the output without much thought. But engineers building production software with AI didn't see themselves as vibe coders.

Simon Willison, the co-creator of Django, recently published a guide that draws the distinction clearly. The moment you begin reviewing and auditing agent-generated code, you've crossed into what he calls agentic engineering.

Pull requests are the first things to break. Willison’s sharpest warning targets a growing anti-pattern where engineers submit PRs full of agent-generated code they never actually checked. This shifts the burden of proof to reviewers. To make matters worse, agent-written descriptions often read well on the surface but fail to match what the code is actually doing.

Experience still compounds if you capture it. The guide argues that every working snippet, proof of concept, and old repository serves as high-quality fuel for agents. Instead of describing a solution from scratch, engineers can give an agent two existing examples and tell it to merge them. This process compounds years of knowledge into a single prompt in a way a beginner simply cannot replicate.

It is cheaper to produce, but not cheaper to maintain. Willison’s core point ties the whole guide together. Agents lower the cost of writing code, but they do not lower the cost of maintaining bad code. Without updated review processes, technical debt grows with every commit.

Ultimately, the cost shifts from the person who wrote the prompt to the team that inherits the repository. If your team is shipping agent-generated code without a clear workflow, Willison’s guide is the place to start.

Claude Code's scheduled tasks used to have a frustrating catch: they only worked if your computer stayed awake. If you closed your laptop, the job would just fail to run.

Anthropic's Claude Code PM, Noah Zweben, shared a newly released fix for this. The “/schedule” command now lets you set up cloud-based recurring tasks directly from your terminal.

Just type “/schedule” followed by what you need:

Claude sets the frequency and gives you a trigger URL. Since the job runs in the cloud, it stays on schedule even if your laptop is closed. Any connected MCP servers like Slack or GitHub work seamlessly within the scheduled prompt.

How Microsoft's AI VP automates everything: This tutorial shows devs how Microsoft’s AI VP uses the AI terminal Warp to automate tedious daily tasks using ad-hoc micro-agents. You'll learn to manage Azure cloud resources via the command line, process files with FFmpeg, and use custom rules and conversational prompts to streamline your workflow.

MiniMax Skills: This repo contains development skills for AI coding agents. By installing them, you provide structured guidance to tools like Cursor, Claude Code, Codex, and OpenCode, allowing them to stop guessing and start delivering production-quality results.

Hyperagents: Current AI stalls because its self-improvement rules are permanently fixed by humans. "Hyperagents" solve this by merging task-solving and self-editing, allowing the AI to continuously upgrade how it learns across any task.

Grow customers & revenue: Join companies like Google, IBM, and Datadog. Showcase your product to our 200K+ engineers and 100K+ followers on socials. Get in touch.

You can also reply directly to this email if you have suggestions, feedback, or questions.

Until next time — The Code team