Welcome back. If this year was all about image and video generation in AI, the coming year is shaping up to be the year of voice generation. Now, Google’s latest release helps developers build powerful new voice agents and voice gen tools with ease.

Also: Creator of Claude Code reveals engineering career lessons, building a functional AI agent with Python, and learn how senior engineers design backends.

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 upgrades Gemini for real-time voice agents: The search giant has introduced Gemini 2.5 Flash Native Audio, a model that integrates audio input, reasoning, tool calls, and speech output into a single unified system. It leads the ComplexFuncBench Audio evaluations with a 71.5% score and achieves 90% instruction adherence. The model also supports live speech-to-speech translation across more than 70 languages. Developers can now build voice agents on Vertex AI or prototype them via the Gemini API preview in Google AI Studio.

Nvidia unveils Nemotron 3 to tackle AI's multi-agent problem: As developers shift from single chatbots to collaborative agent swarms, Nvidia is betting its new open models can help them scale. Nemotron 3 comes in Nano, Super, and Ultra sizes, with a hybrid architecture that cuts inference costs by activating only a sliver of each model's parameters. Nano is available now on Hugging Face, AWS Bedrock, and inference providers like Together AI — and Nvidia is also releasing open training datasets and RL libraries so devs can customize their own agents.

Cursor bridges design and code with new Visual Editor: Frontend work usually means toggling between UI tweaks and searching through files. Cursor’s new browser-based editor closes that gap by allowing you to interact directly with a live web app while an AI agent rewrites the code in real time. You can drag and drop elements, adjust typography and layouts visually, or point to something and prompt text changes. It works on real React apps with near-instant updates.

Creator of Claude Code reveals how he grew his career: In this podcast, Boris Cherny, creator of Claude Code and former Meta Principal Engineer, discusses leveraging "latent demand" and bridging cultural gaps. He emphasizes automating workflows and preparing for an AI-native future where engineers act as orchestrators of coding agents rather than manual implementers.

Two-time CTO reveals what actually drives AI adoption: In this blog, Imprint CTO Will Larson shares hard-won lessons from 18 months of AI adoption at the company. His key insight: if you're not building your own agents, you can't effectively lead the charge. His strategy focuses on removing friction and leading from the front.

Traits of a good tech lead: Expert Head of Engineering João Alves has published a comprehensive guide on what distinguishes great Tech Leads from the rest. His core insight: the best TLs create a multiplier effect within their teams by focusing on architecture, quality, and mentorship — while ensuring they don’t become bottlenecks themselves.

AI Agents in 38 minutes: In this tutorial, you will learn how to engineer production grade AI agents. It guides you through core architectures like reflection and multi-agent collaboration, while covering essential best practices for evaluation, latency optimization, cost management, and security to build reliable, scalable systems.

GPT-5.2 prompting guide (by OpenAI): This official guide will teach you the best prompting practices for OpenAI's GPT-5.2 model — including controlling verbosity, preventing scope drift, handling ambiguity, optimizing tool calling parallelism, and migrating prompts from older models. It also covers compaction for long workflows and structured extraction from PDFs and Office documents.

Building a production grade agent with Python: In this tutorial, you will learn to build a scalable AI web agent using Python, Inngest, and Bright Data. It also covers orchestrating asynchronous workflows, scraping real time web data efficiently, and integrating OpenAI to generate content, while managing concurrency and production level observability.

The state of enterprise AI (by OpenAI): This paper discusses the challenge of turning scattered AI experiments into reliable corporate systems. It finds that moving to structured workflows drives a massive 320x spike in reasoning usage, helping employees save almost an hour every day.

Beyond data filtering (by Anthropic): This paper discusses SGTM, a method that isolates dangerous information into a specific "forget zone" within the AI. Instead of relying on imperfect data filtering, this zone is simply deleted after training, removing the risks while keeping the model's general abilities intact.

SAPO (by Alibaba): This paper discusses SAPO, a solution for the instability that often ruins long AI training runs. By using smoother, "soft-gated" adjustments instead of harsh cut-offs, it keeps the learning process steady and significantly boosts performance on complex tasks.

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

Until next time — The Code team