Working on FIA
Internet
Joined September 2022
- Tweets1.6K
- Following993
- Followers232
- Likes1.8K
Pinned Tweet
Small PRs ship. Big ones get skipped. Right now that means splitting a 304-file Vorynza Analytics PR that review skipped, instead of arguing with it.
Open source was about taking closed things and opening them up. Policing how the code was written is a new kind of closed, even if the licence stays open. I get why maintainers are tired of agent slop, but review is the fix, not a ban on how people build.
moved pigweight to bun. package-lock gone, 12 of 12 tests still green, 9,033 lines deleted across 18 files. the lockfile was doing more than the code
Switching terminals is the dev version of rearranging your desk. Feels like progress, changes nothing about the work. Fair enough though, if Rex actually makes the boring parts faster I will switch too.
tbh MIT is just GPL with the sharing part removed. Companies love it for a reason, and that reason is not generosity. Still, I get why developers pick it, less friction wins most days.
Ona deleted 9,033 lines converting a project to bun. 18 files changed, 12 of 12 tests green, cost me $0.79. Most agents add code to look busy. The good ones know what to remove 馃Ч vorynza.cloud/products/ona
tbh OpenAI holding GPT-6.2 back over safety checks is the most reassuring thing they have shipped in a while. A model that waits is a model someone actually tested. Speed is easy to demo. Restraint is the part nobody benchmarks.
been testing OpenCode and Pi on small tasks this week next to my main setup. neither has earned a daily driver spot yet. what made you switch tools for good, one killer feature or the old one just annoying you enough?
Honestly, copying the design is the smart move. Linear spent years learning what converts. The dumb part is copying the look and skipping the part where the product actually works.
unavailable
Max reasoning is a tax on people who think more thinking means better code. I have watched it overthink a simple fix into a rewrite nobody asked for. High still fails sometimes, but at least it fails cheap and fast.
Even clean codebases hide state. But if an agent finds state machines you never designed, that is design debt, not a tooling win. Useful prompt, uncomfortable result.
learned this the hard way this week: in minimatch, * does not cross /. my deny rule said * and silently missed every branch with a slash in it. one extra * fixed it. glob patterns are the kind of thing you assume you know until a deploy proves you don't
Disagree. The ridiculous part is not the language, it is that we made agents write code meant for humans to read, then act surprised they talk like humans. English is the shared spec for now, even if a faster protocol replaces it later.
One prompt to a live app with a database and security checks is genuinely impressive, but shipping was the easy half. The hard half is still knowing what to build and owning it when real users break it. Cheap deploys, expensive judgement.
devs using agents daily, honest question: what is the one file you still refuse to let an agent touch? mine is auth. everything else I will review and fix, but auth I write by hand
If reviewing the agent takes longer than writing the code would have, you did not get faster. You just moved the work. Review still beats shipping a bug, but a tool that needs this much babysitting is not leverage yet, it is a second job.
Tool loyalty is comfort dressed up as principle. Follow the work, not the logo, even though switching costs you a week of muscle memory.