Welcome back. AI is rewriting the software engineering career path. As agents take over the tactical tasks that junior developers once did, an ex-Vercel engineer warns that we have essentially pulled down the only bridge that brought them into the industry. See his take, plus see what other developers think.

Today: OpenAI quietly dropped a Codex playbook for devs, the breakdown of why Linear feels so fast, and find out where Sam Altman wants OpenAI to head next in AGI.

Tech leaders push devs to fork dependencies: A wave of supply chain attacks on Axios, TanStack, and Nx Console has industry voices rethinking how teams handle packages. HashiCorp founder Mitchell Hashimoto just argued that updating is riskier than leaving latent bugs alone, urging developers to fork dependencies and strip them to the essentials. Other prominent engineers are already backing the move. See the details.

Cohere ships its first fully open-source AI model: The Canadian AI lab just dropped Command A+, a 218B parameter model designed for complex reasoning and agent-based tasks. What makes this release unique is its Apache 2.0 license, which allows teams to fine-tune, deploy, and commercialize the model without any restrictions. It also includes native citations that link claims directly to source data. Plus, thanks to 4-bit quantization, it can now run on a single Nvidia Blackwell B200. Try it here.

Google sunsets Gemini CLI for a new agent terminal: The search giant is moving developers to Antigravity CLI, a Go-built terminal designed for orchestrating multiple AI agents on complex tasks. It handles asynchronous workflows in the background and shares a unified backend with the Antigravity 2.0 desktop app. Existing Agent Skills, Hooks, and Subagents carry over. Free and Pro users have until June 18, while enterprise customers keep access for now.

Foundation models will never see your enterprise codebase. Your proprietary frameworks, internal abstractions, and architectural conventions live exclusively inside your walls. That's why generic AI tools stall on the work that matters most.
Blitzy is built differently:

Start building today.

“Nobody pushed back.” That's the title of a recent post that racked up over 100k views from Hacker News and Reddit this week. It argues that most architectural disasters aren't about missing information. Engineers see what's coming and stay quiet because pushing back costs more than going along. It's an old pattern, and AI is only making it worse.

The shift is mechanical. Software engineering has shifted from writing code to reviewing it. The Pragmatic Engineer's 2026 survey describes developers as "overwhelmed and derailed" by the volume of AI-generated code they now have to audit. The deep intuition that once came from building line-by-line is being replaced by the surface-level task of reading agent output, causing technical depth to thin out.

Dissent depends on depth. Pushing back on a design used to mean knowing the system well enough to spot a failure mode or draft a rollback plan. When that technical depth fades, objections lose their edge. Vague concerns then get rebranded as alignment issues.

Fix the structure. Asking engineers to be braver isn't the answer, especially as AI erodes the deep expertise needed to back an objection. Instead, make blameless postmortems the standard. Require a "risks and unresolved concerns" section in every design document, and create systems that document dissent without forcing a consensus. Harvard's Amy Edmondson covers the AI-era playbook in a HBR piece. Without these mechanisms, silence quickly becomes the default.

Every Cursor session usually ends with a dozen corrections before the agent finally nails your naming and test patterns. But once you close that chat, you're back to square one.

The Atlan engineering team found a shortcut. Just run "/Generate Cursor Rules" after a solid session. Cursor will scan your chat for those specific tweaks and save them into a .cursor/rules/*.mdc file. That way, your next chat starts with your preferences already baked in. Now, pass a description to seed a rule from scratch instead:

Just check the files into Git, and the whole team will pick up the new patterns. The Atlan team suggests running it whenever you find yourself thinking, "ugh, it did that again."

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

Build AI agents for image and video generation: Learn to build AI agents that can generate and automatically evaluate visual media. You will also master advanced evaluation techniques and build two hands-on projects, a UI mockup generator and a multi-scene video explainer, using Google’s models and the Gemini CLI.

Runtime: You can deploy sandboxed coding agents for your entire team. It turns coding agents into teammates anyone can use from Slack, Linear, the CLI, the API or your browser.

Codegraph (11.4k ⭐): A tool to optimize Claude Code by replacing slow file scans with a pre-indexed knowledge graph. Instead of wasting tokens on repetitive tool calls, agents query symbol relationships and call graphs instantly, making codebase exploration significantly faster and more token-efficient.

How to build agents that verify their own work (by Anthropic): This cookbook shows you how to build AI agents that audit their own work. By using a second “grader” agent to verify citations and coverage against a rubric, the system ensures every requirement is met through automated feedback loops.

A single prompt in Claude Code recently uncovered several security vulnerabilities on a founder's MacBook. See if yours is at risk by testing it here.

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