Vercel lets any coding agent drive your desktop

Welcome back. Some devs do their most interesting work on the weekend. This seems to be especially true for OpenAI cofounder Andrej Katrpathy, who built a YouTube clone that serves high signal content. Meanwhile, Vercel announced a 'desktop' for AI agents, and researchers uncovered a flaw in the Linux kernel’s TCP code.

Also: Anthropic’s text watermarks expand, Big Tech’s jaw-dropping quarterly AI spending, and the wild assignment that got a 27-year-old dev unlimited tokens.

Vercel lets any coding agent drive your desktop: The developer platform’s Labs team just open-sourced agent-desktop, a tool that lets coding agents control apps across macOS, Windows, and Linux. Instead of guessing from screenshots, agents read the same app structure that screen readers use, making navigation much more precise. Better yet, it runs in the background without taking over your cursor, so you can keep working. The builder also claims up to 96% fewer tokens on busy apps. Give it a spin here.

Anthropic’s AI turns a Linux bug into root access: The AI lab found a flaw in how Linux handles TCP connections and had Claude turn it into a working exploit. Attackers don’t need a password, and the exploit kicks in before your app sees the connection. It targets servers with a multipath TCP listener. Programs written in Go, Google’s language for building servers, open one by default as of version 1.24. Debian, SUSE, Oracle, and Amazon Linux have shipped fixed kernels but Ubuntu and RHEL still haven’t. If you run Linux servers, here’s what to do now.

Andrej Karpathy builds a YouTube alternative in two hours: The OpenAI co-founder (now at Anthropic) just built HackerTube, a site that surfaces videos the Hacker News community keeps recommending. It sifts through 43M comments to find 6,454 videos worth watching. Inside Claude Code, Opus 5.5 handled the pipeline and deployment, while Sonnet 5.5 did the heavy lifting of labeling comments. Karpathy only stepped in to provide an API key, server, and domain. The labeling reportedly used 504M input tokens, though most were cached.

When data volumes explode, adding another database means more infrastructure, more complexity, and more things that can go wrong.

Tiger Data extends the Postgres you already run:

Start building for free today and get $1,000 in credits.

The 10x illusion. Everyone says AI makes developers 10x faster. But Titus Winters, lead author of Software Engineering at Google, says he hasn’t seen research showing gains above 20% to 30%. Senior engineers tell him code output is up 70%, outages are up 30%, and features shipped are flat. A story from his appearance on a podcast with veteran engineer Gergely Orosz shows what those numbers miss.

One engineer’s hunch. Orosz shared a story about a CTO whose company paid a vendor a revenue share for a proprietary component. Rebuilding it would have taken 5 to 10 engineers years. Then an engineer suggested feeding the specifications and existing test harness into AI. The AI built a working replacement. When the company planned to open-source the project, the vendor offered to buy the code and re-license it. 

Beyond the dashboard. Winters argues that traditional productivity metrics can’t capture breakthroughs like this. His Develop, Deploy, Operate paper offers a better lens. Here’s how you can rethink them.

The hidden cost. That CTO had considered halving his engineering team because AI was making people more efficient. He held off, partly over morale. Had he gone ahead, the engineer behind that windfall might have been let go, and no one would have known what was lost. As senior dev Matt Pocock put it, “Individual contributors have never been more powerful.” Before you cut headcount, consider what your engineers could make possible with AI.

How to build an agentic coding workflow that scales: Cloudroom founder David Ondrej breaks down the setup he uses to ship 10,000+ production commits a year after 15,700 agent sessions. You’ll learn how to combine Opus 5.5, Claude Code, Cloudroom, and reusable skills to spend less time writing code and more time steering agents through the work.

Hex: A local voice dictation tool for Mac, Windows, and Linux that turns speech straight into text wherever you’re working. On Mac, just hold Option, talk, and it pastes your words into the active app, which makes it handy for firing off longer prompts to coding agents without typing everything out.

How to stop overprompting Claude (by Anthropic): This guide shows how to get more out of newer Claude models with less instruction overhead. You’ll learn how to set a clear goal with a verifier, trim rules written for older models, decide when subagents are actually useful, and structure longer-running work with cloud sessions and Projects.

Claude Code takes a half-formed idea and starts writing files immediately. You get working code for the wrong thing. A dev shared a Brainstorming skill that forces a design conversation first.

The skill activates on its own. Claude asks clarifying questions, proposes approaches, weighs trade-offs, and documents the final design before writing a line of code.

Elon Musk shared a Neuralink demo of a speech-restoration patient communicating by thought alone, and the result is wild.

If you enjoyed this newsletter, you should definitely check out our IT and cybersecurity newsletters. Subscribe using the links below:

We put together 20+ hands-on guides covering coding agents, agentic engineering, and leadership frameworks to help you level up your career. You can get them here.

Advertise with us: Join companies like Google, IBM, and Datadog. Showcase your product to our 350K+ engineers and 150K+ 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