Cosmic Rundown: WebGPU Apps, AVIF, and DMARC

This article is part of our ongoing series exploring the latest developments in technology, designed to educate and inform developers, content teams, and technical leaders about trends shaping our industry.

Today's links share a theme. The browser keeps taking on bigger jobs, while the plain mechanics of shipping on the web, such as image formats, email delivery, and CI, keep asking for attention. Below we group the reading into three themes, followed by a set of quick hits.

The Browser Keeps Getting Bigger

KetchUp is an open-source alternative to SketchUp. It is built with Rust and WebGPU and runs as a web app at that link. A 3D modeling tool that opens in a browser tab is a good stress test for WebGPU. The Hacker News discussion covers what it does well and where desktop tools still lead.

The essay The Web Is More Than Enough pairs well with it. The title makes the argument: for most software, the web platform already has what teams need. Readers debated the point on HN.

Two other projects take different angles on the same idea:

• Zebkit packages Chromium as an embeddable library. It is aimed at native apps that want a full browser engine inside them without adopting a whole framework. (discussion)
• Pigshell: Unix the Web applies a shell-style, pipe-everything model to web resources inside the browser. (discussion)

Practical takeaway: WebGPU is moving from demos into real tools. If you ship anything GPU-heavy, test the fallback path on browsers and devices without WebGPU support. Do not assume every visitor has it.

Two Delivery Details That Quietly Break Things

AVIF as the default

Michael Ponders asks Isn't It Time for AVIF Everywhere? The HN thread gets into encode cost, tooling, and how old browsers behave.

You do not have to choose one format for everyone. The usual pattern is to serve AVIF where the browser supports it and fall back otherwise:

<picture>
  <source srcset="hero.avif" type="image/avif">
  <source srcset="hero.webp" type="image/webp">
  <img src="hero.jpg" alt="Product hero" width="1200" height="630">
</picture>

An image CDN can do the same thing through content negotiation. It reads the browser's Accept header and returns the best format the browser supports, so the markup stays simple.

Outlook.com and DMARC

MailRoute reports that Outlook.com is rejecting email from domains without DMARC. This matters for more than marketing email. Password resets, form notifications, and order receipts all come from a domain, and if that domain has no DMARC record, Outlook users may never see them.

A basic setup has three parts:

1. SPF, which lists the servers allowed to send for your domain.
2. DKIM, which signs outgoing mail. Your email provider gives you the DNS record.
3. DMARC, which tells receivers what to do when checks fail. A monitoring-only record is a reasonable starting point: v=DMARC1; p=none; rua=mailto:dmarc@yourdomain.com

Once the reports show your legitimate senders are aligned, tighten the policy. The discussion has useful notes from people who run their own mail.

Agent Tooling: Trust, Memory, and Skills

AI coding tools are now part of daily work, and the questions people ask about them are changing.

• Model drift. A widely shared post asks Is Opus 5.5 nerfed? and claims its thinking patterns changed on November 3. Whether or not the claim holds up, the thread shows a real problem: teams cannot easily tell when a hosted model's behavior shifts. Pinning model versions where you can, and running a small eval set on a schedule, gives you something to compare against.
• Memory portability. Agent memory needs its OpenTelemetry moment argues for an open format for agent memory, so that context is not locked into one vendor. (discussion)
• Shared skills. Jane Street published Agent Skills, a repository of skills for coding agents. It is a useful example of how a large engineering organization packages context for its tools. (discussion)
• Enterprise caution. A report that Goldman Sachs is pulling Claude Code from its bankers and traders set off a long conversation about where agent tools fit inside regulated teams.

Quick Hits

• GitHub Actions incident: Actions had an outage today. This is a good time to check whether your deploy path has a manual fallback. (HN)
• New pull request overview page: GitHub's redesigned PR overview is in public preview. (HN)
• Show HN: DBDelve: an open-source database client built with Rust and GPUI.
• Try Gleam: a case for the type-safe language that runs on the BEAM and compiles to JavaScript. (HN)
• What is malloc doing, anyway?: a readable look at memory allocators. (HN)
• Active Benchmarking: Brendan Gregg's method for checking what a benchmark is really measuring. (HN)
• Pointer Provenance in Rust vs. C++: why pointers are more than integers. (HN)
• H100 rental pricing: the same GPU ranges from $2.69 to $12.29 an hour depending on the provider. (HN)
• Unlocking any car from X brand: a security write-up on connected-car APIs. (HN)

What This Means for Content Teams

Several of today's stories reach editors and marketers, not only engineers.

Images. The AVIF debate is mostly about delivery. Cosmic serves media through an imgix CDN with image optimization and transformations, so adding auto=format to an image URL lets the CDN return a modern format to browsers that support it. Editors upload once, and developers do not have to maintain a picture element for every asset. See the media docs.

Email from your content. If your site sends form notifications, set up DMARC before Outlook users stop receiving them. Our tutorial on handling content form requests with Resend and Cosmic walks through a setup that uses a verified sending domain.

Agents you can check. Model drift is a good reason to keep humans in the loop. Cosmic AI Agents work inside your content model. Workflows add approval gates for human review, along with token usage tracking. If you use coding assistants, the MCP Server and Agent Skills give them structured access to your content and keep that context out of one vendor's memory.

Resilient publishing. When CI goes down, content should not be stuck behind it. Webhooks let you trigger rebuilds from content changes and add retries on your side.

To try any of this, start with the documentation or create a free account.