Welcome back. The tussle between OpenAI’s Codex and Claude Code has taken a new turn. Last week, several developers reported performance degradation issues with Codex, leading many to switch back to Anthropic's Claude Code for more reliable coding support. If you're stuck with Codex and want to migrate to Claude Code — or just want to try it for the first time — this is the perfect tutorial to help you get things done with Claude Code.

Welcome to The Code. This is a 2x weekly email that cuts through the noise to help devs, engineers, and technical leaders find high-signal news, releases, and resources in 5 minutes or less. You can sign up or share this email here.

OpenAI launches a GPT-5 agent that hunts and fixes code bugs: The ChatGPT maker just unveiled Aardvark, a GPT-5-powered autonomous agent that hunts for software vulnerabilities like a human security researcher would. The tool continuously scans code repositories to identify bugs, assess their exploitability, and generate fixes — catching 92% of known vulnerabilities in benchmark tests.

Codex’s performance degradation: Developer communities on X and Reddit are reporting a significant drop in OpenAI's Codex performance, powered by GPT-5, with complaints of slowness and errors leading many to switch back to Anthropic's Claude Code for more reliable coding support. Addressing this issue, OpenAI published “Ghosts in the Codex Machine”, reset Codex rate limits for everyone, and refunded all Codex credit usage up until 1 PM PT Friday.

Moonshot AI releases new open-source faster attention architecture: Chinese AI startup Moonshot has open-sourced Kimi Linear, an alternative to standard transformers that cuts memory needs by up to 75% and speeds up generation by 6× on million-token contexts. Devs can download it from Hugging Face. Here’s a 5-min video to explain how this differs from traditional transformer architecture.

Shopify's distinguished engineer on why being 'the only' beats being 'the best': After turning down a director track at Google to return to engineering, Ilya Grigorik reached Shopify's VP-level distinguished engineer role with an unconventional strategy. His advice for aspiring principal engineers: stop trying to be the best at one thing and start becoming the only person who can bridge multiple domains.

My mistakes and advice leading engineering teams: Gregor Ojstersek's journey through engineering leadership taught him that micromanaging kills momentum and trying to be a hero engineer leads to burnout. His takeaway for new managers: stop trying to control everything, build strong leadership teams, and understand the business deeply.

Take-home exercises for hiring engineers: If you’re looking to hire top engineering talent without getting burned by take-home coding tests, this guide is a must read. It breaks down the three pillars of effective exercises: align with company values, keep scope limited to "an afternoon" of work, and follow a clear process with upfront judging criteria.

How I use Claude Code for real engineering: Developer Matt Pocock just shared his battle-tested approach for using Claude Code on real engineering tasks, demonstrating how to build a complex CLI feature without blowing through context windows. His workflow has already gained traction among engineers looking to scale their AI-assisted development beyond simple scripts.

How to Start a Career and Excel as an AI Engineer: freeCodeCamp just released a comprehensive handbook for aspiring AI engineers looking to break into the fast-growing field. The guide covers everything from must-have skills like Python, machine learning, and transformers to practical career tips on building portfolios and networking strategically.

You don't need MCP at all: Developer Mario Zechner argues that for many use cases, lightweight Bash tools beat heavyweight MCP implementations hands down. His tutorial shows that benefits go beyond efficiency — these tools are composable, easily extensible, and don't suffer from the context bottleneck that plagues MCP servers.

Google’s new AI agent teaches video models to improve themselves: The search giant and NUS introduce VISTA, a self-improving video generation agent that iteratively refines prompts through multi-agent feedback. It improves video quality and alignment by 60% over state-of-the-art baselines.

Alibaba introduces AgentFold, a dynamic memory system for web agents: Current web agents choke on complex tasks by saving too much junk or losing crucial details. AgentFold solves this with smart memory management, beating DeepSeek's 671B model and OpenAI's o4-mini despite being 20x smaller.

Meta drops new AI agent security framework to prevent prompt injection exploits: AI agents that can read your emails, access private data, and take actions on your behalf are prime targets for prompt injection attacks. Meta just released a practical security framework inspired by Chromium's Rule of Two that deterministically reduces these risks by limiting agents to only two of three properties.

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

Until next time — The Code team