Welcome back. If you’re still wondering how to use Claude Code to write production-quality code — we’ve got your back. Check out the Claude Code Playbook, which curates the collective insights of senior devs.

Today: Google launches Computer Use model, senior engineers embrace “vibe engineering”, and learn how to speed up coding by running parallel agents.

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 model turns AI into a power user: The search giant released the Gemini 2.5 Computer Use model, a specialized variant of Gemini 2.5 Pro that powers agents capable of interacting directly with web and mobile user interfaces. It leads on web control benchmarks with lower latency than competing systems.

OpenAI rolls out AgentKit for enterprise AI agents: OpenAI just launched AgentKit, an all-in-one toolkit designed to make building and deploying AI agents faster and easier. It consolidates fragmented tools into a single platform with modules for workflow design, data governance, embedded UIs, evaluation, and fine-tuning. Here’s a 1-hour tutorial to build agents with Agent Builder.

DeepMind’s agent fixes software bugs before hackers can exploit them: Google DeepMind unveiled CodeMender, an AI agent that automatically finds and patches critical security vulnerabilities in software code. It combines reactive and proactive strategies: it rapidly applies fixes for newly surfaced vulnerabilities and proactively rewrites existing code to remove entire classes of flaws.

What are the seven team profiles of engineering delivery performance: Traditional software delivery metrics, such as deployment frequency and change failure rate, tell us what is happening but not why. This practical guide explores patterns of team health beyond traditional software delivery metrics and explains how leaders can diagnose their team's specific challenges.

Senior engineers embrace "vibe engineering": Developer Simon Willison coined the term for the sophisticated, high-stakes use of AI coding agents. Unlike quick-and-dirty vibe coding, this approach demands automated testing, comprehensive documentation, strong code review skills, and the ability to manage multiple agents simultaneously.

How to Build an AI Product Team: Building AI products isn't just about hiring data scientists and hoping for the best. Bryan Bischof, Head of AI at Theory Ventures, laid out a practical framework for assembling high-performing AI teams. His key insight —start lean with a full-stack engineer, data analyst, and product manager before scaling up.

Agentic AI Course: In this course taught by Andrew Ng, you’ll gain a fundamental understanding and practical knowledge to develop production-ready agentic applications, from design patterns to deployment and evaluation.

How AI Agents Cut Cloud Costs by 60%: A platform engineering team deployed five specialized AI agents to autonomously optimize cloud spending, cutting their monthly AWS bill from $380K to $145K. The system hunts zombie resources, rightsizes instances, and manages spot capacity — delivering $2.82M in annual savings with a 58x ROI.

How to speed up coding by running agents in parallel: Explore the “parallel agent lifestyle” to boost productivity without sacrificing code quality. Find out when to launch multiple agents, how to manage risks, and integrate results effectively.

S2R: Google introduces Speech-to-Retrieval that retrieves information directly from audio without converting speech to text first. S2R eliminates transcription errors that plague cascade models, delivering faster and more reliable voice search results for developers.

Agentic Context Engineering: This method helps LLMs improve by updating what they read and remember, instead of changing their core weights. ACE treats the model’s context like a growing notebook that keeps and organizes useful strategies.

Less is More: Recursive Reasoning with Tiny Networks: Samsung's Tiny Recursive Model (TRM) achieves breakthrough reasoning results with only 7M parameters by recursively improving predictions through a two-layer network. It beats models like DeepSeek R1 and Gemini 2.5 Pro on challenging puzzle benchmarks through efficient iterative refinement.

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

Until next time — The Code team