Ibrahim

For fun, I had agents build me a sample app for demoing @Superconductor. 3 out of 4 Claudes built a plant watering app, and 3 out of 4 Codex built excess food distribution apps. The odd ones out were Haiku (coffee tracking) and Spark (neighborhood service). Maybe it's the prompt?

@pierrecomputer with virtualized diffs, do y'all have a recommendation for how to handle "ctrl+f" searching transparently? thanks!

@joemasilotti soloterm.com ? looks neat!

Ruby Native is cooking with Solo today!

A good time to remind people that in my time doing LLM research I feel like a minority of my colleagues are American citizens. It would be industry destroying to have to rebuild with segregation for frontier ai research to be legal.

We are in compliance 🫡

We used coding agents to fix a memory leak in a Ruby gem! I'm interested in @mitchellh's perspective around vendoring the part of dependencies that you use, but there's value in improving the ecosystem so we don't waste tokens validating the same code. x.com/Superconductor…

Fable is a good model. As with all new models, it is simultaneously excellent and entirely unremarkable (relative to other models). It is slow and expensive, and the "loops are all you need" discourse they are pushing is obvious in the context of someone using Fable-class models What I've found so far is that for broad scope design (code architecture) tasks, Fable is unremarkable. Or, not better enough to justify its cost and speed. But in highly targeted goal-oriented loops, it is another beast entirely. It is very slow but produces very good results. I let it churn on optimizing a SwiftUI-layout resolver in Go I wrote and it was able to bring it down to an order of magnitude I could not reach myself (micro => nanosecond scale). But it took 2 hours and $40 to do it and I had to claw back some changes it overfit to Apple Silicon. Still, very worth it. In comparison, for "implement this feature/change" iterative work, I ran head-to-head Fable vs GPT5.5 vs. GLM-5.1. They all produced equally acceptable final results, but GPT5/GLM did it in a couple minutes and Fable was churning away for 40 minutes. And GLM cost me less than a dollar, GPT5.5 ~$1.50, and Fable cost $9. You can see that in this context, interactively working with an agent is nonsense. Its too slow. You need to write loops to keep the agent working and you probably want to highly parallelize the work being done. As with all things, I think a balance makes sense... My sense is that I'd reserve Fable for targeted, surgical analysis and work. Not for daily driving everyday tasks. I'm going to keep spending a shitload of money (relatively) and maining Fable for the rest of the week to continue to judge, will report if anything changes. I'll continue to head-to-head as well.

Good stuff: we launched Fable 5 on a real memory leak ticket alongside GPT-5.5 and Opus 4.8 (a total of 5 agents in one click). Fable was the ONLY ONE that found the root cause: an event listener dependency holding refs to replaced DOM nodes. Red before. Green after. Hell yeah.

github[dot]com

Coding agents let you explore the multiverse of code possibilities. New @Superconductor technique I'm trying: Generate a list of N possible refactors, and then launch N agents instructed to pick the jth refactor and implement it. Allows you to gauge impact of different changes!

@GregMolnar @typecraft_dev Yes. It's totally irrational to want to replace my perfectly fine keyboard but dang does it look good. (Well, more travel and more rigidity would be nice but...)

@typecraft_dev Everytime I see it, it makes me jealous :D

Every time I see this keyboard it just makes me smile.

Time to launch a bunch of @Superconductor implementations and compare results on the train ride home.

@QuinnyPig @BenjDicken Codecommit looks like a summer intern's project that made it to production.

@BenjDicken It's not lost on me that "Codecommit" has been mentioned on Twitter less than two dozen times this month. So uh... AWS does not have the chops here.

AWS could do the funniest thing right now

@joemasilotti You should really check us out at @Superconductor! happy to onboard you personally, your hotwire native content helped us build a pretty sweet iOS app. let me know!

Alright friends, I need help. How in the hell do you set up git worktrees?? I just want to work on two unrelated things across multiple repos at the same time. With different agents. Is that too much to ask? (He says, realizing 6 months ago this would have sounded INSANE).

Running the @simonw pelican vibe-bench is easy in @Superconductor 😅

@antirez Leave the laptop at home and use @Superconductor instead!

When you hit the gym in 2026 leaving your buddy continuing the work.

A subtle benefit of @Superconductor is being able to review code on the go. With @github PR UI I hardly get any context and can't explore the codebase. With agentic code review, I can ask Claude to find similar patterns, explore alternative approaches, and execute code snippets

@nateberkopec mm yeah, that's one reason why running all your agents in cloud sandboxes is safer. you can control exactly what is on the sandbox.

One detail from this story: the AI finding credentials laying around. I have moved to scrub all credentials stored anywhere in plaintext on my system. No more .env, no more ~/.aws/credentials, etc. I use fnox with a 1password backend.

If you want previews in REMOTE claude code sessions, too, check out @Superconductor

@nateberkopec totally agree. but I think the AI PRs can serve as R&D/inspiration for an eventual human implementation. I think there's different levels of fidelity needed for different tasks and the challenge is figuring out what level of human attention is needed

Tobi opened another "AI-powered perf research" PR to bundler in Feb, proposing 37 changes. Today, only 8 are merged, and only 1 of those changes was original to Tobi's PR. The rest were other people's ideas. Further, the merged code is 100% human-written, not Tobi's agentic impl