Welcome back. OpenAI and Anthropic keep finding new ways to make devs more productive. This time, it’s your voice. Both labs just rolled out major voice upgrades to their apps that let you work faster by talking directly as you work — but with two very different directions.

OpenAI and Anthropic race to make voice modes useful: Both rival labs shipped powerful voice upgrades yesterday. ChatGPT’s desktop app now lets you control agents with your voice (available to paid plans). It's powered by the latest GPT-Live voice models to run realtime as you work. Anthropic followed with an update of its own to improve conversational reasoning (see how it works). Claude's voice mode now uses Opus and Sonnet for better reasoning and can connect to your email or calendar mid-chat. 

Notion brings infrastructure-as-code to workspaces: The productivity platform just dropped a beta version that lets teams define teamspaces, databases, and custom agents using TypeScript. Now, engineering teams can review every workspace change to their setup in Git, build setups with coding agents, and recreate tested configurations anywhere without starting from scratch. Apply for the beta here.

Jack Dorsey wants to put AI agents inside your team chat: Twitter co-founder Jack Dorsey's company Block is taking on Slack and GitHub with Buzz. This platform looks a lot like Slack, except AI agents work right alongside humans in the same channels. Developers can manage repos, review code, and pull agents into tasks without ever leaving the app. Buzz is also open source, which means teams can customize their instance to fit their exact workflows. Try it here.

Here's how it works: do the task once with Scribe Capture running. 

It records your clicks and screens, then generates a step-by-step guide with screenshots. 

No writing, no screen recording, no cleanup. New hires follow the guide instead of pinging your senior engineers.

Roll it out to your team

“Graph engineering” is the latest engineering trend gaining traction in Silicon Valley. But behind the label sits a genuinely useful idea: Graphs let you impose your own knowledge of how a system should work, rather than relying entirely on the LLM's judgement.

How they work: Think of a graph as a flowchart your agent must follow. It has two parts: boxes and arrows. A box does one piece of work. It might run some code, ask the LLM a question, or call a tool. An arrow tells the agent which box to go to next. Some arrows are fixed (after step A, always do step B), while others depend on the result (e.g. if the customer's ticket is about billing, go to the billing box, or if it's technical, go to the tech box).

So the whole idea is that instead of hoping that the model picks the right path every time, you make wrong paths impossible using graphs.

Lessons from production:

Graphs or loops? The decision comes down to how predictable your workflow is. If your team knows the steps in advance, encode them as a graph and let the model reason only where it adds value. If the task is open-ended, like deep research where planning and delegation can't be pinned down beforehand, a simple agent loop works better. Read this guide to get started.

How to write better agent skills: You'll learn how to effectively create, structure, and refine custom agent skills. This tutorial covers how to combine markdown instructions with executable scripts and templates, write lean prompts, bridge knowledge gaps, and properly structure global versus project-level configurations.

Fractal by plasmaAI: This tool runs autonomous coding agents inside Git repos. It uses isolated worktrees and a central database to help developers delegate complex tasks to self-steering AI nodes. 

Guide to building and scaling AI agents: Building reliable AI agents shouldn't mean dealing with bloated frameworks or hidden costs. This guide offers a complete breakdown of how to build production-scale agents from the ground up.

AI-written diffs often hit review full of filler comments and formatting noise. This can be a real headache for reviewers. Lee Robinson, an ML engineer at Cursor, shared a pre-PR cleanup routine. It's based on internal tools the Cursor team just went public with. 

P.S. Get 50+ AI coding hacks for Claude Code, Cursor, and Codex here.

Check out this library of 52 prompts to copy into Claude Code. Use it to explore ways of working you haven’t tried or when you’re not sure where to start.

Grow customers & revenue: Join companies like Google, IBM, and Datadog. Showcase your product to our 300K+ 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