Welcome back. A few weeks ago, we talked about how CTOs from billion-dollar giants are ditching their executive titles to become staff engineers at Anthropic. That trend gained its biggest name to date, and the announcement post has already crossed 21M views. See who joined.

Also: Check your GitHub repos for poisoned VS Code extensions, how to audit your Mac for security gaps, and the creator of Claude Code sharing hacks for writing prompts.

Google upgrades Gemini and Antigravity for developers: The search giant just dropped Gemini 3.5 Flash at its annual developer conference, calling it their most powerful model yet for coding and autonomous agents. This model handles coding pipelines end-to-end and even built an entire operating system from scratch during internal testing. Try it via the Gemini API. Google also unveiled Antigravity 2.0 alongside the model, a standalone desktop IDE designed for agent-first development that functions as Flash's native runtime.

Anthropic lets teams self-host agent tools: The AI lab just rolled out self-hosted sandboxes in public beta and MCP tunnels in research preview for Claude Managed Agents during their Code with Claude event in London. This update allows teams to run agent tools within their own infrastructure or through platforms like Cloudflare, Daytona, Modal, and Vercel, while keeping the agent loop on Anthropic's side. Additionally, MCP tunnels let agents reach private APIs without exposing them to the open web.

GitHub admits internal repos were stolen via extension: The code hosting giant just confirmed that a compromised employee's device resulted in the theft of internal repositories. The attacker claims to have grabbed about 3,800 repos, a number that matches the company's own findings. In response, the team removed the extension, isolated the affected device, and rotated all critical credentials overnight. If you have API keys in your code, even private repos, now is the time to double-check and change them.

AI coding tools can speed up writing code, but for most teams, the real bottleneck is finding the right context. This guide breaks down the four categories of AI tools for engineering teams, explains the two-layer model leading organizations are adopting, and offers a practical framework for evaluating tools based on context depth, explainability, security, governance, and workflow fit.

Get the guide to build an AI stack that actually helps your team ship faster and more reliably.

The script keeps repeating. Meta began layoffs in Singapore at 4 AM local time, with 8,000 cuts rolling across different time zones. The internal memo called it a move to create “flatter structures,” echoing Coinbase CEO Brian Armstrong's 14% cut two weeks ago with "no pure managers" left. Now, Airbnb CEO Brian Chesky shared a similar perspective on a recent podcast.

Zuckerberg cut the middleman. He coded his own AI to handle parts of his CEO job like collecting employee feedback, a job once held by managers. Because AI now manages the synthesis and reporting that defined those roles, executives can monitor operations directly, making traditional one-on-ones obsolete. This shift is clear at Airbnb, where AI already writes 60% of the code.

Blame doesn't compile. Some argue that the coordination work being cut was never real management, just admin busywork that leadership should have handled themselves. This creates a paradox where managers are held responsible for results decided by algorithms. For instance, Amazon managers must deal with the fallout of AI-driven quotas they have no power to change, even when those targets push employees to their limits.

Code your way out. If you’re caught in this trap, here’s your playbook:

Companies are cutting titles, but the work itself isn't going anywhere.

When Codex tackles a multi-step bug in a single thread, the debugging clutter piles up and makes the final fix sloppy. Codex documentation refers to this as context pollution.

Simon Willison, the co-creator of Django, shared a fix with OpenAI's custom subagents. Three steps:

Codex runs each agent independently, executing read-only tasks in parallel before consolidating the results.

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

How to optimize context for large codebases in Codex: In this tutorial, you’ll learn to manage Codex's context window using clear and compact commands. It explains how to track token usage, customize the terminal status line, and control conversation threads using fork, resume, and rename functions to optimize AI assistant workflows across large codebases.

Trainer: This tool allows you to train AI agents just by recording your screen. It captures every click, keystroke, and intent as you work, turning your workflow into a reusable agent.

RTK (51.4k ⭐): A Rust-based CLI proxy that compresses command output before it hits your LLM. It slashes token usage by 60-90% for typical dev commands, significantly cutting down on both latency and cost.

Code as Agent Harness: AI agents struggle with complex tasks when they treat code merely as a final output. However, using code as the underlying infrastructure helps these agents plan, test, and coordinate their actions much more reliably.

Grow customers & revenue: Join companies like Google, IBM, and Datadog. Showcase your product to our 290K+ engineers and 160K+ 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