Welcome back. Every AI lab is racing towards the same goal: agents that run in the background, handling complex work across your apps without lifting a finger. OpenAI just shipped a new model that takes us one giant step closer to that future. GPT-5.4 combines reasoning, coding, and professional workflows — and it's their first model that can operate a computer on your behalf.

Also: How to run an entire company with AI agents, the 30 security rules every vibe coder needs, and the ultimate roadmap to becoming an AI engineer.

OpenAI's new flagship model pushes toward agentic coding workflows: The ChatGPT maker just shipped GPT-5.4, its most advanced reasoning model to date. The model features native computer use, allowing agents to write code, execute keyboard and mouse commands, and navigate applications autonomously. It outperforms GPT-5.3-Codex on longer-running tasks, with Codex's /fast mode adding up to 1.5x faster token velocity. See how it works.

Cursor can now monitor and improve your codebase on autopilot: The AI coding tool just launched Automations that let you run agents on schedules or triggered by events from Slack, GitHub, Linear, or any MCP server. These agents run in secure sandboxes, use tools for tasks like editing code or running tests, and learn from past actions. Teams are already using them for security audits, PR triaging, and incident response management. Watch automations in action.

Anthropic pushes back on the Pentagon's supply chain designation: The AI lab is taking the Department of War to court after being labeled a national security supply chain risk. CEO Dario Amodei clarified that the designation is narrow in scope, only prohibiting Claude's use as a direct component of Department of War (DoW) contracts. This leaves the majority of enterprise deployments untouched, though defense contractors should double-check whether their specific integrations are covered by active DoW agreements.

Get exclusive MongoDB Atlas credits, hands-on engineering support, and a unified platform for document data, Vector Search, and built-in scale. 

See how thousands of innovators—representing $200B+ in combined valuation—got from idea to global growth with MongoDB for Startups.

Claim Your MongoDB Atlas Credits

Numbers over noise. Everyone has an opinion on whether AI will replace developers or not, but almost no one has the numbers to back up their claims. Anthropic just published a study using real usage patterns, and for the first time, we can quantify the impact.

Top of the exposure list. Computer programmers rank as the single most exposed profession, with a 74.5% task coverage rate. The tasks being automated aren't abstract — they're writing, updating, and maintaining software: the core of the job.

The gap that matters. LLMs could theoretically handle 94% of computer and math tasks, but real world usage sits at just 33%. That 61-point gap is what Anthropic calls the number to watch. As it closes, employment effects may follow.

The front door is quietly closing. While mass layoffs have yet to materialize, hiring for workers aged 22 to 25 in high-exposure roles has dropped by 14% since the launch of ChatGPT. Rather than resorting to layoffs, companies are simply choosing not to fill vacant positions for now.

Every command executed by Claude Code dumps its entire output directly into the context window. A brief 30 minute session can easily hit 150K tokens, most of which is just noise like passing tests, progress bars, and verbose logs.

To solve this, a developer created rtk, a CLI proxy that compresses command outputs before they hit the context window. To get started, make sure you have Rust installed and run:

Then wire it into Claude Code with one command:

This hook intercepts every command Claude executes and automatically compresses the output. Sessions that previously peaked at 150K tokens are now reduced to ~45K. You can run “rtk gain” to view your total savings.

How to orchestrate coding agents: This tutorial shows developers how to use the AI-powered Warp terminal and its cloud feature, Oz. You'll learn how to fix bugs, navigate local code diffs, and use APIs to trigger automated cloud agents that resolve GitHub issues directly from your workflow.

Awesome-Selfhosted: This is an ultimate anti-subscription playbook with a collection of open-source network services and web applications you can host on your own servers.

Can AI hide its chain of thought (by OpenAI)? Researchers worry advanced AI might hide its internal reasoning to evade safety monitors. Fortunately, current tests show that models still struggle to manipulate their own thought processes, making them easier to monitor and keep secure.

Grow customers & revenue: Join companies like Google, IBM, and Datadog. Showcase your product to our 200K+ engineers and 100K+ 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