Archive
- Anthropic rewrote its Messages API in Rust with no Rust devs
Claude wrote the Rust, humans wrote the plan and diffed old against new on real traffic, and CPU dropped 40 to 50%. swyx confirms the story is real.
- Thorsten Ball: you don't need branches either
Ball followed his case against pull requests with a harder one, push everything straight to main, and says his 21-person SOC2 team already does. Mario Zechner meanwhile took parallel agent runs from 5 per second to 439 by batching database commits.
- Utah gave 34 state devs Claude Code for four weeks
Boris Cherny says 71% of the developers in Utah's four-week Claude Code pilot reported saving 20 or more hours. Meanwhile Thariq's line about technical people thinking their skills are now useless became the argument of the day.
- Malte Ubl: $9 of DeepSeek did a day's worth of work
Malte Ubl says 24 DeepSeek V4.1 Flash sub agents orchestrated by Opus 5.5 cost him about $9, against roughly $2K a day in list-price tokens before. Cheap open weights under a frontier orchestrator is suddenly a real setup, not a thought experiment.
- Theo spent $1,500 in tokens trimming T3 Code's tools
Theo built an 80-case eval from his own rage moments with T3 Code's tools, and Opus failed over half of them. The fix cost $1,500 in tokens and cut tool context by 79%.
- Theo: Gemini hallucinates in over 80% of runs
Theo added a "gemini the jester" slot to his multi-model review skill and says Gemini made things up in over 80% of runs, including a line 173 in a 72-line file. Meanwhile Steinberger's two week agent loop finished a framework migration in about 600 PRs.
- Theo: Agents SDK burns your 5-hour Claude limit faster
A user emptied a full 5-hour Claude Max window in about 30 minutes on one bug-fixing session. Theo spent a day on it and says it is a real bug on Anthropic's side, not your prompting.
- Bun AOT runs Claude Code at 267 MB with 67 subagents
Jarred Sumner spent four hours running Claude Code on a Bun AOT build and watched it sit at 267 MB of RAM with 67 active subagents. He had predicted 2-3x less memory the day before.
- Theo makes three models argue before a merge
Theo now blocks his own merges until Opus 5.5, Muse Wonder Full and GPT-6.1 Sol have argued it out. Elsewhere, a claim that Claude-rewritten Bun has never had security issues got corrected by Bun's own author.
- Theo reverses: agents saved his outage day
Two days after saying his vibe coded pieces collapsed and engineers are still very necessary, Theo says he would not have survived yesterday without agents triaging outages and rolling API keys.
- Steinberger moved a frontend to Solid 2 in ~600 PRs
Peter Steinberger says an agent loop moved OpenClaw's frontend off Lit in about six days and spent another week making it idiomatic Solid, at a cost of roughly 600 pull requests and a few missed button shapes.
- Theo: the vibe-coded pieces collapsed last night
Theo says a night of vibe-coded parts falling over convinced him engineers are still very much necessary, and Kent C. Dodds replied that the thing that saves you is knowledge you cannot quite articulate yet.
- Ball: drop pull requests. Julius: 400 perf PRs in 30h
Thorsten Ball spent the week arguing nobody needs pull requests anymore and published the SOC2 pipeline he pushes straight to main with. Meanwhile Julius let an agent merge almost 400 performance PRs in 30 hours, and Theo says you can feel it.
- 400 perf PRs in 30 hours, with receipts
Julius says his agent merged almost 400 performance PRs in 30 hours and reports 86% fewer renders when another thread streams. Theo's reply says the quiet part: the only reason that is safe is the leak detection someone built first.
- Thorsten Ball: why are you still using pull requests?
Ball says agents make the pull request obsolete and his team pushes straight to main behind feature flags. A colleague's reply: that still sounds like a better PR, not a different thing.
- Pocock: agents that build state machines ship sturdier code
Matt Pocock says he can feel the robustness jump every time an agent reaches for a state machine, and XState's David K agrees agents are now extremely good at writing them. Meta also put an 816B open-weight model into the coding race.
- Theo: Gemini 4 Argon is a liability, keep it off your code
After a day of testing Gemini 4 Argon on his own codebases, Theo says it was factually wrong about more than half its claims in almost every test. That is a different kind of failure than a few points on a bench.
- Thorsten Ball's team pushes to main, no code review
Thorsten Ball says his team ships with no pull requests and no code review, with agents reading production logs and checking their own work. Meanwhile Gemini 4 Argon got a perfect Skatebench-private score from Theo and the worst false claim rate he has measured.
- Theo: Sol's bug fix would not have worked
Theo pointed three models at the same bug. Opus found the root cause, GPT-6.1 Sol wrote 300 lines of tests around a fix he later confirmed was broken, and Muse Wonder Full told him to merge someone else's PR instead.
- Theo's Rust tsc, and the $400k that got nowhere
Five months and about $400k of Codex tokens got Theo nowhere on a Rust rewrite of the TypeScript compiler. About $20k of Opus finished it in two weeks, and he says he has not read a line of it.
- Kent C. Dodds merged 110 PRs without opening his laptop
Kent C. Dodds says he went to bed at 110 merged pull requests in a single day, with six agents still working, and never opened his computer. dax, meanwhile, wants to kill the phrase "rogue agent" before it becomes an excuse.
- dax: one prompt rewrote opencode2, and it cost $500
dax says a single prompt from an iOS app rewrote the whole opencode2 server and TUI, with Opus 5.5 driving subagents for $500. Separately he had opencode rewrite the opentunnel backend in Rust and move it to 10 Fly regions live.
- Thorsten Ball: agents wrote 1,780 lines with no errors
Thorsten Ball is building a compiler one giant single attempt at a time, 1,780 error-free lines in one pass and then aiming at 5,000 to 8,000. He says that is why agents will not need the unit tests humans need.
- Haiku 5.5 swapped in for Luna, and it got worse
ThePrimeagen changed one string in a config file to point his automation at Claude Haiku 5.5 instead of GPT-6 Luna, and says it performed significantly worse and significantly slower. The cheapest model is not automatically the cheap option.