Welcome back. If you’ve noticed some bugs and quality issues when using Claude models lately, you’re not alone. After several weeks of devs complaining, Anthropic has finally admitted what went wrong, and promised a fix that should restore performance.

P.S. We’ve upgraded the design and content of this email after receiving some great feedback from our readers. Let us know what you think in the poll at the bottom.

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.

Google's new protocol lets AI agents handle payments securely: AP2 is a new universal standard that enables AI agents to make purchases on your behalf. The protocol addresses critical trust issues and establishes clear accountability. Dev teams can now integrate agent-driven payments into their apps using one consistent API that works across all merchant platforms.

OpenAI upgrades Codex with specialized GPT-5 model: GPT-5-Codex is an upgraded version of GPT-5 that’s fine-tuned specifically for coding tasks. The new model can work on complex projects for over 7 hours straight, handling everything from quick bug fixes to major refactors.

Alibaba drops first fully open-source web agent to rival OpenAI: Tongyi DeepResearch matches OpenAI's DeepResearch performance across major benchmarks. The 30B-parameter model scored 32.9 on academic reasoning tasks and 75 on user-focused web searches. It works with the ReAct framework without any complex prompt engineering.

IBM launches powerful vision-language model for developers: IBM's new Granite-Docling solves a huge pain point for engineering teams — converting complex documents into machine-readable formats while keeping everything intact, from table structures to code snippets.

Senior Engineer’s Vibe Coding Playbook to Ship Production Features 5x Faster: This guide shows how experienced developers use intuition and pattern recognition to solve complex problems that junior engineers struggle with. Learn how to read codebases quickly, make smart trade-offs under pressure, and guide your team through unclear requirements.

Ex-Stripe CTO reveals the "perfectly happy team" red flag: Singleton warns teams that report perfect happiness rarely perform at their full potential. For engineering leaders struggling with team performance, Singleton has counterintuitive advice: some friction signals growth, while too much comfort might indicate stagnation.

How a non-coding manager became his company's #2 code contributor: Engineering leaders who stopped coding years ago can now contribute directly to production codebases thanks to agentic AI workflows. Skydio's Elliot Graebert scaled from managing tickets to crushing technical debt by running parallel AI agents across multiple workspaces.

Senior devs ship 2.5x more AI code than juniors: A new Fastly survey of 791 developers reveals a surprising split in AI adoption. Senior devs are better at spotting when AI code "looks right but isn't," making them more confident using AI tools for production code.

Google's Official Tutorial for Developers on Agentic Coding: This tutorial teaches how to build, debug, and refactor code directly from the command line using Google's new agentic tool. The CLI features project-aware assistance, multimodal support, and safety sandboxing, making complex coding tasks as simple as conversational prompts.

IBM launches Zero to Production AI Agents Course: This crash course takes developers through four progressive modules, starting with Watsonx's no-code orchestrate platform and advancing to full Python control with LangGraph.

LangChain's Deep Dive Tutorial on Context Engineering: As AI agents become mainstream, developers are hitting a painful wall: context management. LangChain's Lance Martin just outlined 5 key strategies for "context engineering," the art of feeding AI models just the right information at each step.

SFR-DeepResearch: Salesforce researchers have developed a new approach to training AI agents with reinforcement learning. This allows simpler deployment (no complex multi-agent orchestration), better generalization to new tasks, and the ability to handle research problems that require extensive information gathering without hitting context window limits.

Virtual Agent Economies: DeepMind research introduces "sandbox economies" where AI agents autonomously transact using standards like Agent2Agent and Model Context Protocol. Developers must now architect agents that can negotiate resources.

AgentGym-RL: ByteDance’s training framework that teaches AI agents using reinforcement learning instead of copying examples. Results show a 7B model outperformed GPT-4o on multiple benchmarks. For developers, the key insight is that post-training methods and curriculum design matter more than model size alone when building agentic systems.

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

Until next time — The Code team