Ladybird rewrote their JavaScript engine in Rust with AI. Andreas Kling, the main developer, said he ported 25k lines of C++ to safe Rust in about two weeks using AI agents. “We're making a pragmatic decision and adopting Rust as a C++ successor language. What a time to be alive!”

Talking about AI ports: Last week we had the so called Next.js Liberation day. Cloudflare used one engineer and about $1,100 in AI tokens to completely rebuild Next.js on top of Vite. They call it vinext. Almost every line of code was written by AI (Claude via OpenCode).

Does that mean that any open source codebase could be cloned and altered with minimal effort?

TLdraw folks made a joke about that but things are so confusing and happen so fast that people fell for it.

"Sorry folks, this issue was more of a joke (am I allowed to do that?) but I'll keep the issue open since there's some discussion here."

The argument is that essentially, if the agents have access to a good test suite you could write a clone using them as reference. OSS ecosystem is really being tested by this AI wave.

The IDE as we used to know is silently dying. If you're deep into agentic coding you may have realized that we're going to need new abstractions and new tooling for the kind of work we'll be doing in the next decade.

This fits perfectly into Cursor’s latest post about the "Third Era" of AI coding. We moved from Tab autocomplete (Era 1) to synchronous local agents (Era 2), and are now shifting to Cloud Agents. Cursor says over 35% of their internal PRs are generated by autonomous cloud agents.

Still on that front, Stripe built "Minions". An internal end-to-end coding agent. Since Off-the-shelf agents didn't fit the requirements for their massive codebase. They built their own. Anyone who worked at big orgs knows that everything becomes custom at some point. So they built a custom framework that spins up isolated devboxes where agents do the work. They say that over 1,000 PRs merged weekly are completely generated by Minions.

This is another example of the "harness engineering" concept being pushed to the next level. We're at a point that most devs are still figuring out how to get the most out of coding agents like claude code by customizing the many knobs (skills, hooks, subagents, etc), but leading enterprises already started building custom deeply integrated asynchronous fleets of agents. I want to talk more about background async agents in the next issue.

Finally, Android is officially becoming an "agentic" OS. Google launched AppFunctions, which is basically an on-device Model Context Protocol (MCP).

User:

"Find all my receipts for hotels and transport from my trip to NYC last month and move them to my 'Business Expenses' folder in Google Drive.".

AppFunctions flow: This request uses functions from multiple apps. First, the caller uses an email app's search function to identify receipts within the "NYC trip" date range. Then, it extracts the relevant PDFs files and invokes a gdrive app's function to upload them within the specified "Business Expenses" folder.

If this picks up, the way we use phones may drastically change (and a lot of apps might stop making sense).

Gemini already integrates with a bunch of selected apps. Something that I do all the time now is to Launch Gemini and ask it to add whatever I have on the screen (a concert ticket, a flight rebooking email, a WhatsApp message) to my calendar. It magically extracts relevant data and creates an event. Works super well. The same works with Keep, Maps, and much more, but now it will let any app to have this integration.

Local models keep getting incredibly good. Qwen 3.5 benchmarks show the 122B model running at ~57 tokens/sec on an Apple M3 Ultra. That's something unthinkable a few months ago.

If you're unsure what your machine can handle, you can check Alex Jones' llmfit, a great TUI that profiles your hardware and scores models to tell you exactly what fits in your RAM/VRAM.

Pro tips:

For example, if you have a Macbook pro with more than 32gb RAM you can do what I did and get 80 TPS from a very capable model (Qwen3.5:35b-A3B - Gives me Sonnet 3.5 performance vibes) locally!

I had Anthropic’s refusal of the Department of War contract saved to write about, but the fallout is completely insane. Dario Amodei drew a line with the Pentagon: no domestic (f*** everyone else though) mass surveillance and no autonomous killer bots.

The government threatened them with the Defense Production Act and gave a Friday deadline to drop the guardrails. Google and OpenAI employees signed a petition at notdivided.org to block military use of their models.

On Thursday, Sam Altman played the hero and promised OpenAI would stand in solidarity.

Then Friday, the deadline comes and the admin escalates. They label Anthropic a "supply chain risk" to effectively ban any military contractors from doing business with them.

Just a few hours later, Altman magically flips (dude can't help himself). Signs the DoW deal with a contract that starts off on the same premise as anthropic, but basically lets the military rewrite the rules whenever they want. He tried to spin it with no success.

Then the backlash starts. OpenAI's headquarters is covered in protest art in SF and ChatGPT uninstalls just spiked nearly 300%.

Meanwhile, Anthropic becomes a woke, radical left company, as per the current American Admin. Katy Perry tweets her support. And Claude just shot to number one on the App Store.

Come Saturday and the US strikes Iran. After the Pentagon tried to publicly execute Anthropic for insubordination, they still used Claude to run the operation. Harness lock-in is a hell of a drug.

Anyway, this is far from over. But for now we can all sleep well knowing that Altman just plugged a syncopathic Agent Smith directly into the American surveillance apparatus. The NSA golden age is about to start.

llmfit — A terminal tool that profiles your hardware and scores 497 models to tell you exactly which LLMs fit your RAM/VRAM. If you want to run models locally but aren't sure what your machine can handle, this gives you perfect, instant answers.

obsidian-skills — Collection of Obsidian skills/plugins for extending your knowledge base with AI capabilities.

ai-functions — If you're building agentic apps, this is worth a look. A framework for creating structured, callable AI functions.

beautiful-mermaid — This is clever: it converts Mermaid diagrams to ASCII art. The model keeps generating Mermaid syntax, but you get perfectly rendered ASCII output. Hook it up with an agent that edits Mermaid files and you'll get beautiful terminal-friendly diagrams automatically.

voicebox — Voice cloning and synthesis tool for creating custom AI voices.