Welcome back. Chinese open-weight models have caught up to the top US labs, and Washington has taken notice. The rise of Kimi K3 sparked a new push within the administration to ban advanced Chinese models. Recent reports suggest this move could lock in the dominance of OpenAI and Anthropic.

Also: The new hiring process in the agentic era, a guide to scaling agents in production, and a 30-day path to an FDE role.

Your coding agents could be prone to cyber attacks, fix them today: The team at Pillar Security found seven ways to bypass the sandboxes in Cursor, OpenAI's Codex, Google's Gemini CLI and Antigravity. None of them attack the sandbox directly. Instead, the agent writes ordinary config or Git files, and trusted tools on the host machine execute them later. Most of these vulnerabilities have been patched, so be sure to update Cursor and Codex CLI right away.

Linear brings loop engineering to entire teams: The product management platform just dropped Loops. The feature scales the /loop workflows that developers use in Claude Code and Codex across an entire workspace. You can describe a task in plain English. Then, set it to run on a schedule or whenever a specific event occurs. Every run is fully transparent, so anyone on the team can see exactly what the agent did. Give it a try here.

Anthropic makes Claude Code work with screen readers: The AI lab just shipped a screen reader mode for Claude Code. This update swaps the visual terminal for simple, linear text that tools like VoiceOver and NVDA can easily read. Every message now has a spoken label, menus are organized into numbered lists, and a terminal bell rings when Claude needs your input. It brings agentic coding to blind developers. You can enable it right here.

You spend hours on things that aren't code. PR descriptions. Slack threads explaining why you made that architecture call. Linear tickets with enough context so your teammate doesn't ping you at 11pm. Docs you keep pushing to next sprint.

Wispr Flow handles all of it. Speak naturally, and it outputs clean text anywhere you type. Syntax-aware, so your variable names and file paths stay intact.

It won't write your code. But it'll clear out everything blocking you from writing it. Works across Mac, Windows, iPhone, and Android.

Teams at Vercel, Clay, and Rivian already use it daily.

Free to start

The whiteboard flipped. Coinbase candidates can now use AI during every part of the engineering interview. This includes everything from the initial coding screen to system design. Every job offer now even includes an AI fluency rating. The company spent a year rebuilding its hiring process. They just published the full playbook last week.

The job changed first. AI now generates almost all code merged at Coinbase, while humans handle the PR reviews. CEO Brian Armstrong says agents are already running daily loops. These agents ingest feedback, draft fixes, and handle security reviews before handing off to an engineer. Rob Witoff, the Head of Platform, says they rebuilt hiring from the ground up to match this new way of writing software.

The loop got smaller. Two old rounds were 84% correlated. To cut the fluff, the new interview process focuses on three main signals:

Rebuild the loop. Coinbase treated this change like an engineering project. They started with front-end interviews and expanded based on the data they gathered. If your current process still tests for skills that AI has made cheap, this guide from Claude Code creator on building an AI-native engineering team is a great place to start.

Skip the passive lectures. 70% of the 2,200+ sessions are interactive, from workshops and code talks to AWS Jams, GameDays, and the Architecture Rodeo.

Run demos against real workloads with sandbox access in small groups. Las Vegas, Nov 30 to Dec 4. 

Register by August 25 and save $1,200

How Senior Engineers structure large codebases: You’ll learn how to organize scalable SaaS codebases by responsibility instead of just folders. This tutorial teaches you to build a strict six-layer architecture with clear dependency rules to prevent technical debt and keep your project perfectly structured for both engineers and AI coding tools. 

Pinecone: This tool is the go-to vector database for AI agents and apps. It's built for semantic search, retrieving knowledge and long-term memory at scale. 

Cognee (28.8k ⭐): This is an open-source memory platform for AI agents. It gives your agents long-term memory that lasts across every session with a self-hosted knowledge graph engine. 

Self-improvements in modern agentic systems: AI agents should learn and evolve autonomously, yet current research remains fragmented. This study explains how agents improve through two primary methods: updating their internal models or optimizing their operational systems. 

Subagents forget everything between runs because they don't inherit the main session's auto memory. Lydia Hallie is an engineer on the Claude Code team at Anthropic. She shared the fix: adding a memory field to the agent's frontmatter. 

The agent creates its own folder at ".claude/agent-memory/". It loads the MEMORY.md file every time it runs and updates it with new findings. You can commit this folder to share the agent's knowledge with your team.

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

Check out this tutorial from an Anthropic engineer on how to plan, build, and run loops with Claude Code.

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