Google’s new frontier model hunts bugs across codebases

Welcome back. Another day, another Google release. This new model is meant to defend your systems from cyber attacks. Meanwhile, GitHub has unveiled a slew of updates promising a new SDLC, and we can’t wait to try them out.

Also: First impressions of the new MacBook Pro touchscreen, stop Claude from adding itself to your commit history, and a complete guide to all the Grok Bot updates.

Google’s new frontier model hunts bugs across codebases: The search giant just unveiled Gemini 4 Carbon, a model built for long-running engineering work. Google claims the model spent days tracking down and patching bugs across codebases. It also tops Artificial Analysis’ Intelligence Index and more than doubles Argon’s secure-code success rate. Carbon cut successful cheating on Terminal-Bench 4.0 to zero and blocked every attack in Gray Swan’s prompt-injection benchmark. Cyber defenders get access first, and it’ll launch at Argon’s introductory price of $2/$10 per million input/output tokens. Read more here.

GitHub turns Copilot into a multi-repo agent manager: The dev platform wants you managing agents instead of writing every line. Its new Copilot projects run agents across multiple repos at once, tracked in one place. Copilot Fusion picks the best models per task, promising frontier-level results 40% faster at up to 77% lower cost. Microsoft is also bringing its unified Copilot harness to Foundry for building long-running agents. And a new code review tier promises to catch more high-severity bugs. More details here.

Vercel’s new CI keeps pace with your coding agents: The hosting platform just debuted Vercel CI in beta to fix a growing bottleneck. Agents open more PRs than ever, yet most pipelines rerun everything on every push. Vercel CI only runs tasks your changes affect, pulls the rest from cache, and adjusts compute per task. Checks show up in GitHub, so branch protection rules stay intact. It’s available on all plans. Try it here.

Get the web data you need without spending your day figuring out how to access it. Plug the Web Data API into your workflows with just a few clicks.

Claim 10K requests (No credit card required)

AI already has a head start. New engineers used to learn by writing code, fixing bugs, and reviewing PRs. Now AI does much of that for them. Vikram Singh, a Stripe engineer who joined as a new grad about 18 months ago, wrote a guide on getting good anyway. He claims that GPT-6 beats a new hire on day one at understanding the domain, writing code, and reviewing it.

Slow down. AI can find answers, write code, and solve problems so quickly that it’s tempting to let it do all the thinking. Singh warns that while this might make you productive today, it could hold you back in the long run. You risk becoming a “meat proxy” for the model, a term he borrows from Niklas Gruhn. He advises slowing down during those first few months while expectations around your output are still low. 

Understand the code. Singh’s benchmark is HashiCorp co-founder Mitchell Hashimoto’s “whiteboard defense”: engineers should be able to walk someone through any customer-facing system they’ve shipped and justify their decisions. To get there, Singh suggests a few habits:

Know what you’re doing. Singh uses LLMs for most of the “what” questions and saves his colleagues’ time for the “why.” Why was a system built this way? Is this even the right approach? Once you understand the problem, AI can make the actual coding much easier. His bigger point for junior devs and the teams hiring them is that time spent learning early is what makes you genuinely faster later. 

P.S. If you are a junior developer just starting out at your first job, we highly recommend going through this detailed tutorial on Cursor. 

How to make Claude run one task for days: This tutorial shows how to set up long-running agent loops that can keep tackling a hard engineering task for days or even weeks. You’ll learn how to give Claude verifiable goals, use “hill climbing” so it keeps improving its own result, and structure the loop so it doesn’t just burn tokens without making progress.

Reason: An AI teammate built to tackle long-running codebase tasks, run scheduled sessions, and coordinate multiple agents seamlessly. You simply assign the task, intervene when necessary, and let the agents handle the heavy lifting. 

How to build your own Claude Code Mods (by Anthropic): This guide shows you how to customize Claude Code beyond settings and skills with small JavaScript or TypeScript mods that can react to session events, change behavior, or add custom UI. You’ll learn how to get Claude to build one from a simple description, tweak it live, and package it up once it’s useful.

Everyone’s pitching autonomous agent pipelines, but nobody tells you whether your team actually needs one. Senior dev Matt Pocock shared a prompt that settles it, built on his new /software-factory skill which is a glossary for factory design.

Claude interviews you one question at a time against the factory glossary, then brainstorms the infra you’d need to set one up. The skill ships with research docs on orchestrators, git hosts, ticket trackers, and runners.

A Redditor’s wife spent 10 months job hunting, so he built her an AI app. It worked so well he open-sourced it.

If you enjoyed this newsletter, you should definitely check out our IT and cybersecurity newsletters. Subscribe using the links below:

We put together 20+ hands-on guides covering coding agents, agentic engineering, and leadership frameworks to help you level up your career. You can get them here.

Advertise with us: Join companies like Google, IBM, and Datadog. Showcase your product to our 350K+ engineers and 150K+ followers on socials. Get in touch.

You can also reply directly to this email if you have suggestions, feedback, or questions.

Until next time — The Code team