# Swift Newsletter 2026 #1: Agents of change: when Xcode learned to think

# Agents of Change: When Xcode Learned to Think

We've been away for a few weeks. Life, travel, commitments... sometimes they conspire to put the newsletter on hold. But if there's one upside to taking a break, it's that when you come back, you have so much to talk about you don't even know where to begin. And this time, trust us, it's been quite the ride: Apple embracing agentic AI in Xcode, a model war that broke out on the same day at the same time, a Madrid adventure with Apple evangelists that left us speechless, and a drama involving trademark disputes, crypto scammers, and lobsters. Yes, lobsters. Let's dive in.

## Xcode 26.3: The IDE That No Longer Just Helps — It Works for You

On February 3, Apple dropped one of those announcements that redefine how developers work. **Xcode 26.3** ships with built-in **agentic coding**. And we're not talking about some glorified autocomplete here: we're talking about full-blown AI agents capable of searching documentation, exploring your project's file structure, modifying settings, visually verifying results with **Xcode Previews**, iterating through builds, fixing errors, and operating autonomously.

The most interesting part? Apple didn't bet on a single provider. Xcode 26.3 supports both **Anthropic's Claude** and **OpenAI's Codex** as AI engines. And on top of that, it integrates **MCP (Model Context Protocol)**, the open standard that lets you connect any compatible third-party agent or tool. In other words, Apple built the highway and invited everyone to drive on it.

As **Susan Prescott**, VP of Worldwide Developer Relations, put it: *"Agentic coding supercharges productivity and creativity."* And honestly, after trying it out, it's hard to imagine going back. The agents don't just write code — they understand your project, navigate through it, and propose contextual solutions. It's a paradigm shift.

## The Clash of Titans: Same Day, Same Hour

If the Xcode 26.3 announcement shook the Apple world, what came two days later was a full-blown earthquake in the AI landscape. On **February 5**, both **Anthropic** and **OpenAI** launched their flagship models. The same day. At the same time. Well, almost: Anthropic moved its launch up by *fifteen minutes*. Just enough to claim they were first.

**Claude Opus 4.6** (yes, the very model powering this newsletter) arrives with a **1-million-token context window** in beta — five times its predecessor — an **Adaptive Thinking** system that dynamically decides how much reasoning effort to apply based on task complexity, and the headline feature: **Agent Teams**, which lets multiple Claude instances work in parallel while coordinating with each other. To showcase its power, Anthropic built a complete C compiler using a team of parallel Claudes. No human intervention. Let that sink in.

Performance-wise, Opus 4.6 scores **72.7% on OSWorld** (up from 66.3% on Opus 4.5), **65.4% on Terminal-Bench 2.0**, and **80.8% on SWE-bench Verified**. But where it truly shines is in web search improvement (+16 points) and terminal operations (+5.6 points). All at the same price: $5/$25 per million input/output tokens.

Meanwhile, **GPT-5.3-Codex** arrives as a unified model: OpenAI has stopped separating code-specific models from general-purpose ones. It's **25% faster** than its predecessor, supports **400K context tokens** and **128K output tokens**, and here's the mind-bending part: it's the first OpenAI model that *helped create its own version*. Early iterations of the model debugged its own training pipeline. Sam Altman also noted it's the first model to hit "high" on cybersecurity within their preparedness framework — implying both immense potential and significant risk.

## When Apple Opens Its Doors

Now, allow us a moment of pride and gratitude. Because over the past few weeks, we've had the privilege of experiencing something truly special.

On **February 2**, **Julio César Fernández**, **Arturo Rivas**, **Jon Ander Orcera**, and **Alejandro Mohamad**, representing **Apple Coding** and **AC Academy**, attended a talk on **Liquid Glass** at **Impact Hub Madrid**. The session was part of Apple's **"Meet with Apple"** program, the worldwide series of design activities Apple is hosting through March 2026 to help developers adopt their new design language.

The talk was led by **Aniket Roy**, **Design Evangelist at Apple**, who broke down the key principles of the **Liquid Glass** design system — the biggest visual overhaul of Apple's platforms since iOS 7 — and how developers can integrate its principles into their apps. An inspiring session that made clear Apple continues to put design at the very core of the user experience. SwiftUI APIs like `glassEffect`, `GlassEffectContainer`, and `glassEffectID` allow apps using standard components to get Liquid Glass support with minimal effort, but the devil is in the details for custom implementations — which is precisely what these in-person sessions help resolve.

But it didn't stop there. **Julio** also had the opportunity to attend a **two-day technical workshop on the Foundation Model Framework**, led by **Alberto Ricci**, **Technology Evangelist at Apple** since 2008 — a true legend in the field. With a completely hands-on approach, Alberto guided over 20 Apple developers through a real-world programming case that went well beyond expectations: not just the **Foundation Model Framework** itself, but also **real-time transcription**, **image generation**, **text-recognition scanning**, and many more on-device AI capabilities. Two intense days that proved Apple's on-device AI is no toy — it's a serious, powerful tool ready for production.

We want to sincerely thank **Annikki Laine**, **Partnership Manager for Worldwide Developer Relations, Northern Europe at Apple**, who from Helsinki makes it possible for these activities to reach developers across Europe, and the entire Apple Developer Relations team for opening their doors to us. Moments like these remind us why we love this community.

## OpenClaw: The Lobster That Set the Internet on Fire

Speaking of community, it's impossible to ignore the phenomenon that has shaken the tech world these past weeks: **OpenClaw**. But to understand the full story, we need to rewind a bit.

It all started with **Clawdbot**, a "weekend project" by **Peter Steinberger** — yes, the founder of **PSPDFKit**, one of the most recognized names in the iOS ecosystem who sold his company after a $116 million investment from Insight Partners. The idea was equal parts simple and ambitious: an autonomous AI agent, open source, running on your own machine and connecting to your everyday apps. **WhatsApp, Slack, Discord, iMessage, Telegram, Signal, email, calendar, terminal**... Clawdbot could manage it all. In an interview with The Pragmatic Engineer, Steinberger dropped the line that went viral: *"I ship code I don't read."* In January 2026 alone, he made over **6,600 commits**. Just him. From his home.

The name "Clawd" was an obvious nod to Anthropic's Claude. And of course, Anthropic soon came knocking with a trademark complaint. Steinberger asked if he could simply drop the "d" and go with "Clawbot." The answer: *"Not allowed."* So he renamed it to **Moltbot**. And here's where it gets wild: in the **ten seconds** between releasing the old GitHub and Twitter handles and registering the new ones, **crypto scammers snatched both old accounts**. Absolute chaos ensued.

After 48 hours of confusion under the Moltbot name (which "never quite rolled off the tongue," Steinberger admitted), **OpenClaw** was born — with a lobster as its mascot and the determination that this would be the final name.

The numbers are staggering: from 9,000 to **157,000 GitHub stars in 60 days**. **34,168 stars in just 48 hours**. It's the fastest-growing repository in GitHub history, 18 times faster than Kubernetes. It even spawned **Moltbook**, a social network *for AI bots* where users' agents posted content, commented, and voted on each other's posts. It reached **1.5 million registered agents**. Andrej Karpathy (former AI director at Tesla) called it *"the closest thing to sci-fi I've seen in a long time."*

The capabilities are real and impressive: agents that prepare personalized morning briefings, email and calendar management via text message, full project scaffolding, pull request reviews from your phone, WhatsApp group monitoring with facial recognition... The list of what the community is building with OpenClaw is endless, with hundreds of *skills* on its **ClawHub** marketplace.

But — and this is a very important but — **we're using it and we know firsthand that you need to be extremely careful**. A critical vulnerability has been discovered (**CVE-2026-25253**, CVSS score 8.8) enabling one-click remote code execution. Security researchers found **341 malicious skills** on ClawHub — 335 from a single campaign. **Snyk** analyzed nearly 4,000 skills and found that 7.1% contained critical flaws exposing credentials. Prompt injection attacks can hijack your agent through something as innocuous as a Google Doc. And there's no sandboxing by default.

OpenClaw is an incredibly powerful tool. But if you're not someone with solid technical experience who understands the risks of giving an AI agent access to your digital life, it can become a security nightmare. The Register called it a *"dumpster fire"* for security. XDA published an article titled *"Please stop using OpenClaw."* Use it, but with great, great respect. And if you do, review every skill you install, enable sandboxing, and don't grant permissions you don't understand.

## Swift Conquers New Territories

While the AI world was on fire, Swift kept marching forward on multiple fronts.

### Windows Is Now Swift Territory

On **January 26**, the Swift project officially announced the **Windows Workgroup**, a dedicated working group to ensure ongoing Swift support on Windows. The goal: improve the official distribution, adapt core packages like **Foundation** and **Dispatch** to Windows conventions, and establish best practices for bridging Swift with the native Windows API. Swift is dead serious about going cross-platform.

### Swift 6.3 Is on the Horizon

**Swift 6.3** continues its path toward a spring 2026 release, with significant **Embedded Swift** improvements: floating-point printing, the `@c` attribute ([SE-0495](https://github.com/swiftlang/swift-evolution/blob/main/proposals/0495-c-attribute.md)) for C-compatible functions, improved LLDB debugging, core dump analysis, and the `@section`, `@used`, and `@export` attributes. Additionally, the **Android Workgroup** is making progress with daily snapshot builds, and there's preview support for **FreeBSD 14.3+**. The language keeps expanding into territories that would have seemed like science fiction just a few years ago.

### Evolution in Motion

Several **Swift Evolution** proposals are under active review:

- **[SE-0506](https://github.com/swiftlang/swift-evolution/blob/main/proposals/0506-advanced-observation-tracking.md)**: **Advanced Observation Tracking**, enhancing `@Observable` types with more fine-grained control over property change detection. Especially useful for middleware infrastructure and widget systems.
- **[SE-0507](https://forums.swift.org/t/se-0507-borrow-and-mutate-accessors/84376)**: **Borrow and Mutate Accessors**, enabling direct property access without copy overhead. A significant step for performance.
- **[SE-0504](https://github.com/swiftlang/swift-evolution/blob/main/proposals/0504-task-cancellation-shields.md)**: **Task Cancellation Shields**, introducing a mechanism to "protect" cleanup code from task cancellation. Authored by Konrad Malawski.

### Making C Libraries Feel at Home in Swift

A must-read [post by Doug Gregor](https://www.swift.org/blog/improving-usability-of-c-libraries-in-swift/) on swift.org explains how to make C libraries more ergonomic in Swift using **API notes**: YAML files layered onto C headers to produce native Swift enums, automatic memory management with `SWIFT_SHARED_REFERENCE`, idiomatic function signatures with `SWIFT_NAME`, and type safety with `OptionSet`. All without touching a single line of original C code. It even includes a Swift script to auto-generate the API notes. Brilliant.

### Your iPhone Thinks, but Doesn't Spy on You

Apple's **Foundation Models Framework** continues to gain traction among developers. We're talking about a **3-billion-parameter model** running entirely on-device: no data sent to the cloud, no additional costs, working even offline. With as few as three lines of Swift code, you can access summarization, entity extraction, creative text generation, short dialog, and much more. Apple has published new [tutorials on Foundation Model prompting](https://developer.apple.com/documentation/FoundationModels) and an [official code-along](https://developer.apple.com/videos/play/meet-with-apple/205/) that's very much worth your time.

### Community Radar

- The **Swift Student Challenge 2026** is accepting submissions from **February 6 through 28**. Distinguished Winners will be invited to Cupertino this summer.
- On **February 10**, Apple is hosting a full day of **foundational SwiftUI sessions** from Cupertino with a livestream.
- **Studioworks** shares its success story building complete web applications in Swift with **Hummingbird 2**, **Soto**, and **Elementary**, achieving sub-100ms page load times.
- Notable new packages: **Cadova** (programmable 3D CAD tool), **Feather Database** (database-agnostic layer with Swift concurrency), and **MailFoundation** (port of .NET mail libraries).

### And That's a Wrap

What a few weeks it's been. Apple embracing agentic AI in Xcode, two tech giants launching models on the same day, a lobster wreaking havoc on GitHub, and us having the privilege of learning firsthand from Apple evangelists in Madrid. If one thing is clear, it's that 2026 isn't going to let up. And honestly, we wouldn't have it any other way.

If you'd like to drop us a comment or share anything, we're on the usual channels.

**See you soon, and until then: cheers and Good Apple Coding.**

Thanks for reading our newsletter! See you on Be Native.