Welcome back. Every productive developer knows the pain of switching between sessions on Claude Code. Well, not anymore: Anthropic just solved it for us. Meanwhile, a new story reveals how hackers used Claude to breach OpenAI.
Also: Building cool weekend projects with Jev (the viral AI model), the mystery behind Union Alpha revealed, and a crash course in agentic evals from Hugging Face.
Today’s Brief
An inside look at OpenAI’s agentic software factory
GPT Astra deciphered a 1918 German radio message
How to build with the Jev model (tutorial)
Cleaning an AI-generated codebase (cookbook)
TODAY IN PROGRAMMING
Click here to watch Claude Code’s Projects in action.
You no longer have to juggle sessions on Claude Code: Anthropic just dropped Projects in beta and rebuilt Claude Code around it. You can now brief Claude like a chief of staff all in one thread. It then hands off tasks to parallel cloud threads that open PRs, run tests, and have shared context. Each thread can also break down its own tasks using subagents, loops, and workflows. Watch it in action or grab creator Boris Cherny's actual prompts.
Devs are scrambling to get their hands on this new AI model: ChatGPT co-inventor Diogo Almeida’s newly launched model, Jev, came out of stealth this week, and early users are already wiring it into everything. They’re using Jev to build model routers, review PRs, and even made a context compaction tool that can slash a Claude Code session from 1M tokens to 86K in a second. Jev is probably on every engineer’s social feed this week but the access is limited. You can get on the waitlist here.
Hackers used Anthropic's Claude to breach OpenAI: Security researchers at Hacktron found a bug in “libheif,” the image library used by OpenAI’s help forum, and turned it into something much bigger. A rigged image let them run code on the server remotely, while a separate login flaw let them hijack employee ChatGPT and Codex accounts. Since those accounts were connected to other tools, that opened the door to internal GitHub code, Slack, and email. They used Opus 5 (reportedly with loosened cyber guardrails) to pull off the exploit. See the breakdown.
PRESENTED BY AGENTFIELD
Instead of juggling a terminal for every agent, you get one window for every project.
Hand work off: it splits it into parts, runs them in parallel, tests what comes back and merges what passes.
It asks only when it must.
Built for open models like DeepSeek, Qwen, GLM and Kimi, or any provider you choose.
Every task keeps a page with its brief, tool calls, report and cost.
One small Go binary, Apache 2.0. Everything stays in plain files on your disk.
INSIGHT
An inside look at how OpenAI is building an agentic software factory
Source: The Code, Superhuman
The idea. What if shipping software looked less like writing code and more like running a factory full of agents? That’s the agentic software factory veteran engineer Gergely Orosz says OpenAI is building internally. In this, Codex agents can take a change from idea to production, and humans mainly set direction and only step in when something looks risky.
Think of it as the software version of a “dark factory,” where automation handles most of the floor work. Here’s the setup:
Codex builds. A developer describes the change. Codex reads the codebase, Slack, and internal data, writes the code, runs tests, fixes failures, and opens a PR. It keeps going until CI turns green, while another tool checks for performance regressions before merge.
Agents review and ship. Multiple agents inspect the PR, each focused on areas like cloud, infrastructure, or security. Low-risk changes can be approved automatically. After human sign-off, a deploy agent ships the code, creates monitoring dashboards, and can roll it back if something breaks.
The factory watches itself. Perf Factory, a service that monitors production, sifts through alerts, removes duplicates, and traces slowdowns to their cause. During incidents, Sevbot, an incident-response agent built on Codex, gathers context and suggests fixes without applying them. The goal is a system that can handle routine releases and incidents, pulling humans in only when judgment matters.
The catch is cost. This setup burns through huge amounts of compute and tokens, something OpenAI can absorb more easily than most teams. Orosz thinks that gap will shrink as token prices fall, making agentic software factories eventually affordable for everyday engineering teams, not just frontier labs.
P.S. If you’re looking to adopt a software factory model in your team, check out this guide from the founder of Warp.
PRESENTED BY PRIORLABS
Your model took weeks to build. Testing TabPFN against it takes seconds.
Run it on the same problem before you ship.
If your model wins, you know. If TabPFN wins, you know before putting it into production.
IN THE KNOW
What’s trending on socials and headlines
Meme of the day.
Agent School: Want to train your own coding agent? This six-session series by HuggingFace covers evals, fine-tuning, and RL, and every line of code is public.
Code Cracked: GPT-6 Astra deciphered a WW1 German radio cipher that sat unsolved for over a century and then proved its answer with a clever trick (1.4M views).
The New Lazy: Shopify’s CEO has a name for one of the worst AI habits creeping into teams, and once you hear it, you’ll start spotting it everywhere (2.5K likes).
Mystery Solved: Union Alpha, the stealth model everyone was testing yesterday, just pulled off its mask. Turns out it's not one but several models working as one (315K views).
Ship It, Defend It: AI makes it easy to push code you do not understand. A veteran founder shared a simple test to determine what you can responsibly ship (7.8K likes).
Ad Erasure: A dev used TypeSafe AI's new model, Jev, to build a realtime adblocker extension. No filter lists involved, and it claims to be undetectable (3.5K likes).
TOP & TRENDING RESOURCES
Click here to watch the tutorial.
Top Tutorial
How to build with Jev: Unlike standard LLMs that generate text, Jev scores possible outcomes in milliseconds. This tutorial shows you how to set up the API and build three projects: a voice-controlled browser, an AI memory system, and a YouTube idea scorer. It also covers when you still need a standard LLM to back it up.
Top Tool
Terminal-browser (3k ⭐): Open a browser directly inside Claude Code, so you can inspect sites, follow links, and work with web pages without bouncing between your terminal and a separate browser window.
Trending Cookbook
How to de-slop an AI-generated codebase: Telling coding agents to "clean this up" never works because it is too vague. In this guide you’ll learn to turn recurring messes into concrete checks agents can actually execute. You’ll see how to use lint rules to catch bad patterns, write strict acceptance conditions so agents do not just silence warnings, and spot when a local fix demands a full architectural refactor.
AI CODING HACK
How to make Codex threads work together
An engineer on OpenAI's Codex team shared a fix for a common Codex pain: parallel threads that know nothing about each other, forcing you to copy-paste context between them.
Threads can now reference, message, and monitor each other. Here's the setup.
Reference another thread from any composer. Type
@and pick the thread, or drag it in from the sidebar:
Read @Fix-route-distance-and-turns and apply the same fix here.Turn one thread into a coordinator that messages the others:
Monitor the "Watch navigation" and "Lock Screen" threads.
When both finish, review their diffs and send each thread
any fixes needed.Add a schedule to the coordinator thread and it checks the fleet on its own. You set up the work and review results instead of driving every thread by hand.
P.S. Get 50+ AI coding hacks for Claude Code, Cursor, and Codex here.
IN CASE YOU MISSED IT
Our most-clicked story from yesterday
Don't throw away your old Apple devices. A dev just showed what $30 of dead hardware can become with Claude and some clever hacking
Grow customers & revenue: Join companies like Google, IBM, and Datadog. Showcase your product to our 350K+ engineers and 150K+ followers on socials. Get in touch.
Whenever you're ready to dive deeper
We put together a few guides on coding agents, agentic engineering, and leadership frameworks to help you level up in your career. Browse all our guides.
What did you think of today's newsletter?
You can also reply directly to this email if you have suggestions, feedback, or questions.
Until next time — The Code team
