A daily pick of what is changing in tools, models and agents for people who build websites. Original summaries, reviewed by a person.
Code Layers moves into open beta: start from a repository, explore variations with the Figma agent, turn them into editable layers and push changes to a GitHub branch. Another step towards design and code living in the same place.
Gemini 4 Argon is Google’s new top model, aimed at multi-step work such as coding and agentic tasks. It is available now through the Google AI Studio API at an introductory $2/$10 per million input/output tokens, doubling after November 30. Google says that inside Antigravity it increased agent code submissions and cut rollbacks.
With Modern Web Guidance installed, the coding agent builds declarative dialogs with invoker commands instead of boilerplate, while keeping cross-browser support in mind. Part of a weekly Chrome series on the topic.
Syntax shows an agent that starts on its own from a Sentry error, can be paused and resumed, and ends up opening a pull request. It runs on Pi Durable, a library for agents that save their state and survive restarts.
midudev turns six photos of a restaurant into a landing page where scrolling walks you through the venue, using Claude Code and the Higgsfield MCP to generate the clips. It shows the terminal workflow and the trick for stitching segments together. Sponsored by Higgsfield.
Figma lets you add guideline files to a library so its agent follows your design system conventions without restating them in every prompt. The video shows how to set it up once for the whole team.
All three coding agents write everything they do to disk, in undocumented formats. The author of Agent Standup, an open-source tool that reads those files to show which agent is working or waiting, explains where they live, which records matter and how sub-agents show up. Useful for anyone monitoring several sessions at once.
A developer hid 40 bugs in a test online store and had Claude Code explore it with Playwright using three prompt styles. With a full checklist it found about 30 bugs on average for under a dollar per run; five bugs were missed by every run. The takeaway: a useful first pass before release, not a replacement for human testers.
EmDash is an open-source CMS from Cloudflare built on Astro and designed for both humans and agents. This tutorial shows how to install it outside Cloudflare, on a Hetzner VPS with Node.js, SQLite, Caddy and PM2, making it an option worth weighing against WordPress for Astro projects.
Meta has released the weights of Muse Spark Open, a multimodal model with about 816B parameters (42B active) and a one-million-token context, aimed at coding and long agentic tasks. It is on Hugging Face and available through providers such as Together AI, Fireworks and OpenRouter, although its license restricts commercial use.
Vercel recaps its annual event, focused on infrastructure for agents: AI SDK 7 for building multi-turn, tool-using agents, MCP Events in beta to react to MCP server events, and new budget and privacy controls in AI Gateway. On the traditional side, Python becomes generally available and Daemons arrive for background workloads.
Void is an open-source SDK, CLI and platform that turns a Vite app into a full-stack Cloudflare Workers app, inferring and provisioning resources such as databases, queues and AI bindings. It integrates with Astro, React Router, SvelteKit and Nuxt, among others, and its CLI is designed so agents can debug, read logs and deploy.
A pilot between Cloudflare and OpenAI uses signals from Cloudflare’s network so ChatGPT search can find new pages, know when they change and skip wasted requests. It matters to site owners because it shapes how AI search discovers their content; Cloudflare says it relies on metadata, not page content for training, and access is still managed with AI Crawl Control.
The author of a directory of AI tools for developers explains how rankings are modeled as Zod-validated data in Astro content collections. The post also covers making the site readable by chatbots and letting an AI agent ship changes without bypassing the rules.
Vercel’s monthly report on AI Gateway traffic shows that in September open-weight models handled 71.5% of tokens but only 23% of spend, with DeepSeek leading on volume. It is useful context when picking production models: price per token is falling, but total spend keeps growing.
Vercel Drives decouples an agent’s files from the environment it runs in, so it can switch sandboxes and keep the same workspace. It is now available for Vercel Sandbox on all plans, and drives can be mounted by several sandboxes or forked to try alternatives in parallel.
The new release adds a --timings flag to see where build time goes, separates build and runtime image services and adds a notFound() handler for custom entrypoints. Parallel prerendering and custom redirect pages arrive as experimental features, useful for large static sites.
The new Gutenberg plugin release loads React 19 unless an incompatible theme or plugin prevents it, which is worth testing in custom builds. Notes also move to a margin beside the content, Navigation block link search is unified and more blocks gain design controls.
