Whether AI coding assistants actually increase developer productivity is a contested topic right now. I believe they do. In my experience, the output gains are real and significant. What used to take days can now be generated in minutes. But with this newfound speed comes a dangerous side effect: complacency. Developers are becoming less critical of the code they produce, and even less critical when reviewing code generated by their teammates' AI tools. The result is that bugs, performance issues, and technical debt are slipping through at an alarming rate. The solution is not to abandon AI. It is to double down on the one thing AI cannot replace: a strong pull request review culture.

The 'Does It Work' Trap

When a developer prompts an AI agent to generate code, they naturally shift into a specific mindset. They focus on whether the code works, not how it works. The AI produces hundreds or even thousands of lines of code in seconds, and the developer runs it, sees it function, and moves on.

This is a trap. Functional code is not the same as good code. I have seen AI generate solutions that introduce multiple full table scans in a single API request. The code worked. It returned the correct data. But in production, with real data volumes, it would have caused devastating performance issues for users. The developer who prompted the agent did not catch this. A reviewer did.

The pace of AI-powered development tools has picked up dramatically. It feels like every month brings a new editor, model, or agent promising to change the way we build software. Some push boundaries. Others stumble. All of them are shaping how developers think about code in this new AI-augmented era.

In this post, we’ll walk through what’s happening across the landscape and take a closer look at Kiro, a new entrant that brings some surprisingly thoughtful ideas to the table.

VS Code Tries To Catch Up

Microsoft recently rolled out an AI Agents update that finally puts agents directly inside the editor. The catch? Using it feels like waiting in a never-moving security line. You can't selectively approve a batch of commands; instead you must click yes for every single tool invocation it dreams up. After the tenth pop-up your flow is less pair programming and more whack-a-mole approval training.

Cursor Trips Over Its Own Pricing

I’ve been thinking about this for a while now, and after a few in-office discussions with a coworker, I’m becoming increasingly convinced: Rust might just be the backend language for the AI era.

That’s a bold claim. But here’s my take.

Rust: Annoyingly Strict, Gloriously Reliable

If you've ever written Rust, you know what I mean when I say it's not exactly a cuddly language. The compiler doesn’t hold your hand. It slaps it away and tells you to rewrite your function signature. And your lifetimes. And your error handling. And your soul.

But here's the magic part: if your code compiles, it's very likely to just... work.

A couple of months ago, I started experimenting with using Gemini as a pull request reviewer in our primary frontend repository. It cannot approve pull requests on its own, but it provides suggestions and considerations for developers to act on.

The initial feedback from the team was overwhelmingly positive. Our frontend developers found the comments helpful and relevant, and after reviewing some of them myself, I decided to take the next step: enabling Gemini by default across all pull requests in all repositories.

Where the Magic Happened

This is when things got interesting.

As part of our yearly security obligations, we undergo external penetration testing. One of the findings this year pointed out that our OneTimePassword generator used the default Random class in C#, which is not cryptographically secure. The fix was simple, switch to a secure random generator.

The last few months have completely reshaped how I think about building software.

From Curiosity to Productivity

It started with small experiments in Visual Studio Code using agent mode. At first, I saw it as a fun way to automate routine coding tasks. But once I learned how to write effective prompts and create detailed playbooks, things started to change. The results got a whole lot more interesting.

Discovering Better Tools

That led me to Cursor. Their AI integration felt smoother and more intuitive than what I had seen in VS Code. My productivity improved. Not dramatically, but noticeably. Then came a downside: a rogue background agent ignored spending limits and racked up over $300 in a single night. Not ideal.