#022 • 27 May 2025

Contents of this Newsletter

🦾

Code writes itself

⏰ : 5 min

📓

Recent Blog

⏰ : 10 min

🗿

New AI Companion

⏰ : 3 min

🗞️

Weekly AI News

⏰ : 7 min

MAIN STORY

The Week Code Wrote Itself, And What That Means for QA

If it felt like the entire dev world jumped ahead five years last week, you’re not alone.

In the past week, we saw a collection of AI agent launches across the board. OpenAI rolled out their Codex agent—an early look at an AI that can not only understand tasks, but ship code end-to-end. Google announced "Jules," a code assistant built to operate in the background, catching bugs and suggesting changes before you even realize you need them. At Microsoft Build 2025, Microsoft CEO Satya Nadella, announced a lot of dev focused updates, starting with making GitHub Copilot in VS Code open-source, and announced that Copilot is evolving from autocomplete to a fully-fledged AI coding agent. And Anthropic dropped Claude Opus 4, called it the best coding model available today, and openly stated they’re stepping back from the AI chatbot race to focus on building the future of dev tooling.

That’s not a trend now. That’s a shift in how software is being built.

And while all of this is exciting—if not a little overwhelming, it made us pause. All this progress, but one question kept coming up:

Why is code generation advancing faster than nearly every other AI use case?

Why Is Code AI's Favourite Playground?

There are a few reasons. Some are technical. Others are cultural.

First, code is structured and self-validating. Unlike natural language, where meaning is often fuzzy, code has a compiler or runtime to tell the model whether it's right or wrong. That feedback loop is highly important for learning.

Second, developers write everything down. Stack Overflow, open-source repos, documentation—all of it adds up to a massive training ground for LLMs. It’s not just a lot of data, it’s the right kind: well-organized, domain-specific, and outcome-oriented.

And third, developers are more willing than most to adopt rough, early tools—especially if those tools shave off even a bit of grunt work. Coders have been debugging their own code with AI long before the rest of the world figured out how to prompt.

It makes sense, then, that this is where we’re seeing the most aggressive acceleration.

A Quick Roadmap Of How We Got Here

Here’s a quick look at the AI coding arc so far:

Early Years (2021–2024)

2021: GitHub Copilot launched, powered by OpenAI’s Codex, offering simple autocomplete-style code suggestions.
2022: GitHub Copilot became generally available, expanding access to AI coding tools.
2023: GPT-4 introduced multi-file understanding, while specialized models like Meta’s Code Llama emerged.
2024: Anthropic’s Claude 3 and Google’s Gemini models advanced coding with deeper reasoning, setting the stage for more autonomous tools.

The Rise of Vibecoding (Late 2024–Early 2025)

Cursor: An AI-first editor with agent mode, gaining major traction.
Lovable: A no-code tool for building apps from natural language prompts.
Others: Replit’s AI agent and tools like Bolt embrace the same natural-language-to-code vision.

Agent Era (2025 So Far)

February 24, 2025: Anthropic released Claude Code in a limited research preview, alongside Claude 3.7 Sonnet, enabling developers to delegate tasks via the terminal.
May 16, 2025: OpenAI’s new Codex, an AI coding agent integrated into ChatGPT, launched, handling multiple tasks simultaneously.
May 19, 2025: GitHub Copilot’s coding agent entered public preview, enhancing autonomous coding workflows.
May 21, 2025: Google’s Jules, an asynchronous coding agent, was released into beta, allowing developers to assign tasks and review results later.
May 22, 2025: Anthropic’s Claude Opus 4 was released, advancing AI coding with deeper reasoning and agentic capabilities.

Throughout 2025, Cursor and Lovable continued to be prominent, reflecting the ongoing rise of vibecoding tools. What used to be AI writing code is quickly becoming AI delivering features.

Where QA fits into all this

When code gets written this fast by agents who don’t necessarily wait for you to review everything, quality becomes even more critical.

The speed of code ≠ the speed of confidence
AI agents can ship code in minutes. QA pipelines, historically, are not built to move that fast. Manual QA becomes the bottleneck. Even traditional test automation can’t keep up when PRs are generated, tested, and merged in the same hour.

That’s why we believe QA has to shift left—really left. At Quash, we enable teams to move toward true test-driven development by generating test cases from PRDs and Figma files, even before a single line of code is written. It means teams can define and validate what "correct" looks like before the first commit.
More code = more surface area for bugs
If AI triples a team’s output, it doesn’t necessarily triple their ability to check for regressions or side effects. Especially when agents operate without the full context of business logic or legacy quirks. We’ve designed Quash to do more than run test scripts. The goal is to track changes, monitor flows, and flag suspicious behavior across the system.
QA is Moving From Correctness to Context
In a world where AI agents write and ship code, traditional QA checklists aren’t enough. Developers won’t just ask “Did the tests pass?” They’ll ask: “Can we trust what the agent just shipped?” That’s a shift from outcome validation to system observability.

QA is slowly becoming the radar system, constantly scanning for intent drift, behavior shifts, and early signs of risk. Not just direct hits like bugs, but subtle signals that something’s off. Agent-generated code might pass every static test and still miss the point. Imagine a requirement that says: “Redirect expired users to the pricing page.” The agent does that but skips showing the warning banner or waiting until the user accesses premium content. It followed the spec, but missed the experience.

This is why Quash is built to stay in sync with your real sources of context, including the code itself. As your system evolves, Quash dynamically generates and updates test cases to reflect actual implementation changes even those that weren’t captured in a PRD or a Figma file. So you're not testing outdated assumptions or static suites. You're validating what the system is right now and not just what it was meant to be. QA is shifting from binary checks to semantic alignment: not just does it work, but is this what we intended, in context.

Final Thought

We’re entering an era where entire features are specified, coded, tested, and deployed without a human typing a line. It’s wild and it's fast and that's a reality we will all get used to. And at Quash, we’re not just testing software—we’re testing the shift in how software gets made.

So if the code is writing itself, QA is about keeping the system accountable. We’ll keep building toward that future. One test case ahead of the code.

Check out our Recent Blogs

Innovation

It’s The Time For Design Led AI Hardware

Ben Geskin

@BenGeskin

WSJ on Jony Ive and Sam Altman’s OpenAI device: The product will be capable of being fully aware of a user's surroundings and life, will be unobtrusive, able to rest in one's pocket or on one's desk, and will be a third core device a person would put on a desk after a MacBook

9:18 PM • May 22, 2025

763

Retweets

8839

Likes

Read 1027 replies

Imagine a device smaller than your phone. No big screen, no endless notifications. Just quiet intelligence, always aware, gently guiding. Now imagine it designed by Jony Ive (the Apple design legend) and powered by Sam Altman’s AI vision.

Recently, they have joined hands to launch a new AI hardware venture called io. The news sparked a wave of excitement and skepticism online. Some people are intrigued by the idea of a sleek, screen-light AI companion that blends seamlessly into everyday life. Others wonder, why add another gadget when our phones already do so much?

There’s also a deeper unease: a device that’s always listening and watching raises privacy concerns that feel eerily close to dystopian visions like George Orwell’s 1984. It echoes the chilling tone of Black Mirror, a show known for introducing eerily plausible gadgets that blur the line between innovation and dystopia. Is this the next step toward constant surveillance, or a clever way to reduce screen time and help us live more mindfully?

The truth is, no one knows yet. But this collaboration signals that AI hardware won’t just be about tech anymore; design and intention will shape how these devices fit into our lives. As we get closer to this future, it’s worth asking: what kind of AI companion do we really want?

news

This Week in AI

Major AI Breakthroughs at Google I/O 2025

Google I/O, Google’s annual developer conference held May 20-21, 2025, in Mountain View, CA, unveiled over 20 AI-powered innovations, reinforcing Google’s dominance in AI development. The event showcased cutting-edge tools and platforms for developers and businesses. Key highlights include:

Gemini 2.5 Advancements: The Gemini 2.5 Pro and Flash models feature enhanced reasoning, native audio output, and robust security. Gemini 2.5 Pro’s Deep Think mode leads benchmarks like LMArena in complex math and coding tasks.
AI-Enhanced Search: Now available to all U.S. users, AI Mode supports complex queries, integrates with Gmail for personalized results, and enables features like virtual clothing try-ons and custom sports/finance charts.
Creative Tools: Imagen 4 offers advanced text-to-image generation with improved detail and text rendering. Veo 3, paired with the Flow app, generates AI-driven videos with synced audio for filmmakers and creators.
Project Astra: A multimodal AI assistant that uses phone cameras for real-time tasks like research and app navigation, integrated into Android XR glasses with partners like Samsung and Warby Parker.
Jules Coding Assistant: A public beta AI tool for autonomous coding, tackling bug fixes and feature development, competing with Microsoft and OpenAI solutions.
Google Beam: Previously Project Starline, this AI-powered 3D video communication platform, launching with HP, enhances virtual meetings with real-time translation.
Subscription Plans: AI Ultra ($249.99/month) provides exclusive access to tools like Veo 3 and Flow, while AI Pro ($19.99/month) offers broader AI feature access.

Claude 4 is here: Anthropic goes all-in on AI coding

Anthropic has launched Claude 4, with a clear pivot from chatbot competition to dominating the developer tooling space. The new models—Claude Opus 4 (premium) and Claude Sonnet 4 (faster, more affordable)—are built for long-horizon coding tasks, parallel tool integration, and multi-step reasoning.
Key features:

Long-memory task execution for complex workflows
Real-time integration with dev tools
Sustained focus across tens of minutes or more
Designed to support agent-based development

Anthropic claims Claude Opus 4 is the best coding model on the market today, signaling a bold shift toward becoming the go-to platform for AI-assisted software engineering.

Sam Altman & Jony Ive Announce ‘io’—a new design-led hardware venture with OpenAI

Sam Altman and Jony Ive formally unveiled io, a new company born from their two-year collaboration to rethink how humans interact with AI.
What is io?

A hardware-focused company founded by Jony Ive, with key Apple veterans (Scott Cannon, Evans Hankey, Tang Tan)
Built in partnership with OpenAI to create a new class of AI-native products
Merging now with OpenAI for tighter integration with research and engineering

Why it matters:
This marks a bold step to move beyond traditional screens and interfaces—designing new tools that let people interact with AI in more natural, human ways. Jony Ive and his studio, LoveFrom, will now lead creative direction across both io and OpenAI.

All Newsletter Issues | Security | Contact Us | Blogs | Test Case Generator

AI Agents Shipping Code—Can QA Keep Up? + Altman & Ive’s io 🔥