Welcome back. Web layouts have been constrained by CSS for decades, limiting how we interact with text. One Midjourney engineer had enough and built a tiny TypeScript library that handles layout 500x faster. The demos went viral within hours.

Also: How to harness long-running AI agents, OpenAI dropped a repo of ready-to-use Codex prompts, and how Claude found unknown security flaws on its own in a live demo.

New TypeScript library makes web text render 500x faster: A Midjourney engineer just open-sourced Pretext, a pure TypeScript engine designed to measure and position multiline text without relying on CSS or DOM reflow. Developed over several weeks using Claude Code and Codex, it validates against browser ground truth to support all languages, RTL scripts, and emojis. The repository quickly gained traction with 11.9k GitHub stars as demos went viral.

Mistral drops a lightweight text-to-speech model: The French AI lab shipped Voxtral TTS, a 4B-parameter model that delivers realistic, expressive speech in nine languages. It can clone any voice from just a three-second clip, capturing natural nuances like pauses, accents, and intonation. In human evaluations, it outperformed ElevenLabs Flash v2.5 on naturalness with latency of just 70ms.

OpenAI's Codex adds plugins to streamline dev workflows: The ChatGPT maker just dropped plugins that bundle MCP servers, app integrations, and reusable workflows into simple one-click installs. There are over 20 plugins available at launch, including tools like Figma, Linear, Notion, Sentry, Slack, and Gmail. This brings Codex in line with Anthropic's Claude Code and Google's Gemini CLI. All three major AI coding agents now support installable plugin packages.

Launching a startup is all about moving fast, but too many startups incorrectly scale their agent fleets, wasting precious resources and eroding trust.

Google Cloud’s startup agent guide shows you how to nail your agent fleet the first time, so you can deploy without worry. 

Master your agent journey here.

Frontier labs got outscored. Intercom CEO Eoghan McCabe announced last week that their custom model, Apex 1.0, is now outperforming GPT-5.4 and Opus 4.5 in customer service. It resolves 73.1% of issues, hallucinates 65% less, and responds faster than the general-purpose giants. With Fin already handling 2 million issues a week, Intercom is nearing $100M in revenue.

Cursor played the same card. Just days later, the coding startup shipped Composer 2. This model, fine-tuned on the open-source Kimi K2.5, beats Claude Opus 4.6 on terminal benchmarks at just one-tenth the cost. Both companies followed the same playbook: take an open-weight base model and post-train it on proprietary domain data until it beats frontier models.

Data is the moat. Intercom built Apex using billions of real-world customer service interactions curated by its 60-person AI team. Cursor trained on millions of actual coding sessions. This domain-specific data is something frontier labs simply don't have. This creates a massive gap between SaaS companies that can train vertical models and those that can't.

Labs face a squeeze. Andrej Karpathy recently noted that AI is shifting toward specialization instead of one-size-fits-all models. Frontier models are currently over-serving specific industries with general intelligence that simply costs too much. Companies with direct access to real-world data are already building cheaper, faster alternatives, forcing the big labs to either specialize or watch their margins disappear.

Every time Cursor tries to run a test or build your project, it pauses and asks for permission. You click "allow," it runs, hits another snag, and asks all over again. YOLO mode solves this by letting the agent execute commands autonomously, but simply toggling the setting isn't enough. You need a scoped prompt to whitelist safe commands.

Go to Cursor Settings > Features and enable YOLO mode. Once active, you'll find a prompt field that dictates which commands the agent can run without intervention. Paste this:

Now, when Cursor hits a TypeScript error, it automatically runs “tsc”, fixes the issue, and restarts the build until it passes. You can swap in your own tools (like “pytest” or “cargo build”) and use the deny list to block destructive commands like “rm -rf” or database migrations.

How to build a deep research agent: This tutorial teaches developers how to build a deep research AI agent from scratch using the Google Gemini API. You will learn to implement core agentic workflows, including tool calling, state management, and deploying sub-agents to handle complex, multi-step tasks efficiently.

Background Agents: A coding agent system that runs tasks in full development environments via Modal. It offers fast-start sandboxes, multiplayer sessions, and PR creation using your own GitHub or AI providers like Claude and OpenAI, making it ideal for self-hosted or internal team use.

Agentic AI and the next intelligence explosion (by Google): Many expect future AI to be a single, massive super-brain. However, researchers discovered that advanced models solve hard problems through internal debate, proving the next AI leap will actually be a collaborative, social network of humans and machines.

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