Welcome back. China just set a new boundary in the AI arms race, and it starts right at its own borders. Top AI researchers at Alibaba and DeepSeek are now required to get government clearance before traveling abroad. See why.

Also: How to build the LLM stack from scratch, Claude Code settings every dev should know, and why Cursor's VP wants engineers to keep thinking for themselves.

Google unveils its coding CLI and no-code Android builder: The search giant shipped Antigravity CLI, a lightweight terminal interface that runs the same agent harness as Antigravity 2.0 while syncing with your custom setup. Google’s AI Studio also got an upgrade that lets anyone build native Android apps, with users creating over 250,000 apps in the first week since launch.

xAI opens its terminal coding agent to subscribers: The Musk-led AI lab just dropped an early beta of Grok Build for SuperGrok and X Premium Plus subscribers. This terminal CLI gives engineers the power to review or rewrite an agent's plan before execution and hand off bigger tasks to parallel subagents in their own worktrees. It automatically picks up existing AGENTS.md files, plugins, and MCP servers. It also comes with a headless mode for scripting. See how it works.

OpenAI turns Codex into a Mac cleanup tool: The ChatGPT maker just found an unlikely second job for Codex, its macOS coding agent. After a developer asked it to run a read-only Mac audit, the agent flagged up to 500GB of reclaimable space sitting in Docker leftovers, Xcode simulator data, and old AI model files. The thread blew up after OpenAI President Greg Brockman shared it, with other devs chiming in about their own 200GB+ storage wins.

Your dev stack got an AI upgrade everywhere except the input layer. You're still typing every prompt, every ticket, every review comment by hand.

Wispr Flow closes that gap. Dictate into Cursor, VS Code, Slack, Linear, or anywhere else you work. It's syntax-aware: camelCase, snake_case, acronyms, and file names all come through clean. Mention a file in Cursor or Windsurf, and it auto-tags.

It's the voice layer for an AI-native workflow. Speak your intent. Your tools do the rest.

Available on Mac, Windows, iPhone, and Android. Used by millions of developers, including teams at OpenAI and Mercury.

Try free

Agents are writing half of Cursor’s code. Cursor revealed this week that over 40% of their codebase’s pull requests now come from cloud agents, a massive jump from just a tiny fraction six months ago. The Cursor team shared their findings, noting that most breakthroughs happened in the infrastructure supporting the model rather than the model itself.

The environment is the bottleneck. A cloud agent starts with a blank slate. Unlike a local agent that inherits a laptop’s shell, dependencies, and network access for free, a cloud agent needs all of that rebuilt from scratch. When a dependency is missing, the agent doesn't just crash; it silently produces lower-quality work, leaving you to blame the model. Cursor ran into this so often they essentially built "enterprise IT" specifically for agents.

Scaling for reliability. As agents began running for hours instead of minutes, simple server hiccups started killing entire workflows. Cursor’s initial setup failed frequently enough to become a problem, so they rebuilt their system on Temporal, a workflow engine designed to survive machine failures. Today, it reliably handles over 50 million actions every day.

But the harness keeps shrinking. Interestingly, as models get smarter, they need less oversight. Early versions of Cursor required constant hand-holding and complex hardcoding, but the latest version just gives the agent a CLI and gets out of the way. This shifts the engineering focus toward building better infrastructure rather than managing model behavior. For teams building cloud agents, Cursor's full writeup is a great reference for seeing the architecture and logic behind these shifts.

Picking the right model for a coding task usually means trying Sonnet, deciding it missed something, switching to GPT, rewriting the prompt, and then giving Composer a shot. Cursor shipped a fix to run all three at once in isolated Git worktrees.

Open “Agent Chat” in the classic editor layout instead of the new Agents Window. Make sure you are inside a Git repo so the worktrees function correctly. To start, run the command using a comma-separated list of models followed by your specific task.

Cursor spins up a worktree for each model and runs them all at once. Once they're done, you can compare the results side by side and pull the best one into your main workspace by using: "/apply-worktree."

P.S. Get 50+ AI coding hacks for Claude Code, Cursor, and Codex here.

How to use Devin for agentic coding: This tutorial teaches developers how to set up and use Devin, an autonomous coding assistant. You'll learn to leverage the local CLI for your daily build, spin up cloud agents for complex tasks, run parallel sub-agents, and streamline workflows with GitHub, Slack, and Linear integrations.

Unabyss: A personal context layer that syncs your identity, knowledge, and preferences across all AI tools. It pulls from your apps to give every agent instant, accurate context so you can stop copy-pasting and start shipping.

Anthropic Cybersecurity Skills (9.6K ⭐): This repo gives your AI agents professional cybersecurity skills. It provides every agent with senior-level expertise for threat hunting, incident response, and building secure apps from scratch.

Memory as a Model: AI models get outdated fast and struggle to pick up new info without expensive retraining. To solve this, researchers built a framework that stores new knowledge in a separate memory model that any frozen AI can easily pull from.

Claude Code creator Boris Cherny is endorsing a course that helps developers with zero terminal experience start shipping real apps.

Grow customers & revenue: Join companies like Google, IBM, and Datadog. Showcase your product to our 290K+ 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