The record you leave behind is training data now

The week's better work was about leaving a readable record of a system. Andrew Kelley, who created the Zig programming language, published a response to Bun's Rust rewrite that reads less like a rebuttal than an audit: Bun is a JavaScript runtime, it was built in Zig, it has been rewritten in Rust, and here is what the move actually bought. Elsewhere a team wrote up moving a production agent to a newer model as a behavior diff rather than an upgrade note, and an onboarding CLI parses a repository into the map it never shipped with. We wrote a couple of weeks ago about the intake path, the part of a system that decides what a model reads before it answers. That path runs backwards too, into what teams publish about themselves, because the public writing around a project is where agents learn to use it. Careless example code teaches a model bad habits, and the model hands them right back. The tell will be whether anyone rewrites the examples, not the front page.

Top Stories#

Bun's Rust rewrite got a public accounting, and it skipped the original reason. Bun is a JavaScript runtime that was written in Zig and has now been rewritten in Rust. Andrew Kelley, who created Zig, went through that rewrite from the Zig side, and the notable part is an omission: compile times, the complaint that motivated the whole move, go unaddressed while every other outcome gets its celebratory paragraph. Anyway, most of what a rewrite delivers comes from building the thing correctly the second time, with the problem finally understood, and that gain is usually available without changing languages. People credit the new language because that is the change they can see. Meanwhile, the genre keeps growing, because building in the open has become how developer infrastructure markets itself: here is what we built, here is what worked, here is what did not.

Migrating a production agent is porting work now. Ploy, a company running an agent in production, published notes from moving it to GPT-5.6, which is the sort of document nobody writes until model upgrades stop being drop-in. The behavior differences are worth planning around. On low-level work the newer model tends to reach for stack traces and a debugger where the older one would sprinkle in log lines and rerun, which is better and also slower: a model that investigates spends wall-clock time that a model that guesses does not. Sure, that is a fine trade at 2am on a real bug, and a worse one on a queue of small chores. Teams that pinned a model version for stability now have to say whether they were buying capability or a fixed set of habits.

Codebase onboarding is being treated as an indexing problem. Onboard-CLI parses a repository into a syntax tree and builds the map locally, which is to say it reads the code as structure instead of shipping the whole tree at a model and hoping the context window holds. The premise is that a new arrival, human or agent, needs structure before prose. It also points somewhere less comfortable for tool authors: how well an agent works with a project is largely downstream of what other people wrote about that project in public, and a local index is the part a team can fix without waiting on the next pre-training run. Finding the file is the easy half, and getting back the piece that actually answers the question is the part still worth working on.

Top Stories#

Top Links#