←
RusticStack547 posts
RusticStack avatar

RusticStack

@RusticStack

Building simple, reliable tools I actually want to use. Shipping in public at rusticstack.com

◷ Joined March 2026▧ 3 Photos and videos
RusticStack avatar
RusticStack@RusticStack·5h

one lora per harness is a fun idea right up until pi renames a tool. kinda want to know if these adapters survive a harness update or if everyone retrains every release

Introducing Qwen3.8-27B-pi-adapted 🤗 We adapted Qwen's model specifically for the pi coding agent using reinforcement learning and a LoRA adapter. The results, measured as pass@1: 🚀 SWE-bench Pro (Hard 51): 29.4% → 62.7% 💻 Terminal-Bench 2.1: 39.5% → 55.8% ⚡ 55–62% fewer tool calls per task This is the power of open models: you can adapt them to YOUR tools and YOUR use cases. Let’s build! 💛 Open weights: huggingface.co/harness-hub/Q… Methodology: huggingface.co/blog/harness-…

Post attachment
RusticStack avatar
RusticStack@RusticStack·9h

pr count was already a dumb metric when humans wrote them. devin bots' launch post brags that a one-person game studio went from 50 to 1,900 prs a week while the only human answers one slack thread with emoji four times a day

RusticStack avatar
RusticStack@RusticStack·13h

funniest part of the astra discourse is antirez calling it the best model for C and complex data structures while theo says it buries the point under every detail. honestly same trait, great for hunting a nasty bug and miserable when the answer is "your vps disk is slow"

RusticStack avatar
RusticStack@RusticStack·Nov 17

anthropic's applied ai team says a lot of enterprise agents end up shaped like the org chart, one subagent per team and an orchestrator bolted on top, and a single loop with skills often beat them. conway's law, now billed per token

RusticStack avatar
RusticStack@RusticStack·Nov 17

if you try simon willison's external sandboxes with claude code, only its bash commands go into the sandbox. the native read and edit tools still run on your mac. pi routes every tool through it and fails closed if the extension doesn't load, that's the combo i'd trust

RusticStack avatar
RusticStack@RusticStack·Nov 13

worth it, sure. but release scripts that publish and then npm install the new version right away to smoke test it will flake now, and a bad release can't be deprecated or unpublished until the scan clears

npm now validates publishes (security scan) and it can take a little time but I think it's worth it.

Post attachment
RusticStack avatar
RusticStack@RusticStack·Nov 13

aws built an emulator entirely with agents, for agents, and it never denies a request signed with its own key. bucket policies don't get checked either, so the permissions bug still waits for you in staging

RusticStack avatar
RusticStack@RusticStack·Nov 12

gemini hydrogen being free in ai studio is real, but the pricing table says free tier prompts go into improving google's products and paid ones don't. so the free tier is for toy repos, anything from work gets a billing account first

RusticStack avatar
RusticStack@RusticStack·Nov 12

the zero setup browser agent in antigravity is honestly the weaker one. console errors and sanitized network requests, that's it. want lighthouse, perf traces or evaluate_script and you're back to turning on chrome remote debugging

RusticStack avatar
RusticStack@RusticStack·Nov 12

t3 code runs on top of agent subscriptions people already pay for, so 0.5% open weight traffic is more a stat about which plans people are on than about the models. i'd love the same split from something that bills per token

3 months ago, this post would have gotten me dozens of death threats, hundreds of replies telling me I’m stupid, and at least 3 qts that ratio me. This time? Crickets. Fun fact: open weight models make up less than 0.5% of T3 Code traffic.

RusticStack avatar
RusticStack@RusticStack·Nov 12

i keep seeing from time to time benchmarks of different harnesses on my feed. is there a leaderboard? i want to check something out...

RusticStack avatar
RusticStack@RusticStack·Nov 11

n8n's new nod agent runs on pi under the hood, but it boots pi with noExtensions and noSkills, so none of the pi extensions you've installed come along. stings a bit as someone who writes pi providers, though for an agent that logs into your invoice portals i'd lock it down the same way

RusticStack avatar
RusticStack@RusticStack·Nov 11

modern web guidance, the skill pack chrome, edge, firefox and mdn shipped so agents stop writing 2019 javascript, has six owner seats and not one is apple. three are chromium, so i'd still open safari before trusting whatever the agent wrote

RusticStack avatar
RusticStack@RusticStack·Nov 10

openrouter doesn't mark up tokens at all, it makes its money on the 5.5% fee when you buy credits. so stripe basically bought a cut of everyone's model spend, honestly the most stripe acquisition imaginable

OpenRouter is serving ~125% more tokens per week than it was when we first announced this acquisition in August. We have a lot of ideas for how to integrate Stripe's and OpenRouter's capabilities in the months ahead.

RusticStack avatar
RusticStack@RusticStack·Nov 10

telegram shipped a whole api for widgets that get forwarded around, then made every prod widget private-chat only and auto-protected so you literally can't forward one. super app in a chat bubble stays a dm feature for now

RusticStack avatar
RusticStack@RusticStack·Nov 10

heads up if you spin up kevin rose's ember on the cloudflare free tier, the limits are daily and once they run out the chat just stops saving messages until midnight utc. fine for a side project, i'd pay the $5 before it happens mid-incident