Claude Code sometimes says "done" when its own tests failed. We open-sourced a fix: one Stop hook, two-minute setup. On our agent it cut replies needing a correction by 40%.
609.sol retweeted
the ai dopamine loop is delayed pretty cooked. too much waiting.
spending most of the day waiting for ai should not be the norm.
if companies want to see real ROI on their AI spend they can’t wait for fully autonomous agents to do people’s jobs.
i’m pretty convicted that low latency, cost-efficient, high intelligence models are going to kick off the real age of ai productivity.
fast, cheap, smart.
if employee at hyperscaler works from 9-5, ideally their productivity goes up 10x from all the ai spend
if they’re spending most of the day waiting to see the results of their work, it’s no surprise they aren’t doing 10x the work.
but if ai delivers frontier intelligence at 1/10th the speed, they can get into flow state and actually gauge deadlines for quality work
this also protects the american frontier labs’ business model as i believe they will be able to optimize/serve their models at fastest speed/intelligence/cost via API compared to open models
trying to figure out best ways to deploy capital against this thesis, just sharing tho
Found a bug running @openclaw in prod for @VibeKitBot: agents were quietly losing their memory summaries on OpenRouter routes.
Fixed it at the source, wrote a live repro to prove it, and @steipete just merged it into main.
Officially a 🦞 contributor.
every agentic coding demo shows the writer. in production, the writer is the cheap half. the verifier is the product, and the loop has to close on production truth, not the model grading its own homework. what running thousands of agent loops taught me: vibekit.bot/blog/loop-engine…
WE ARE LIVE!!!
Introducing Minds by @animocabrands, the AI agent for you and the people around you to share.
It remembers, it connects, it simplifies. So you don't have to do it all alone.
An AI with your best interest at heart.
Built VibeKit — an iOS app + Telegram bot that ships full-stack apps from your phone.
How I built it: I was running an OpenClaw agent on an EC2 box that also hosted one of my projects. Drove the agent from Discord on my phone — chat in, code committed, deployed in place. Realized that workflow should be a product.
Architecture: Node + Express gateway on a single EC2, per-user AWS Fargate containers (one per app) each holding the user's repo + an OpenClaw agent, nginx routes <subdomain>.vibekit.bot to the right Fargate task, Supabase for users/apps/agent_messages/billing, Stripe for payments.
BYOK Anthropic/OpenAI routes direct (no markup), otherwise OpenRouter (Qwen free, others +20%). iOS is pure SwiftUI native. Telegram bot uses the same backend and talks to the same agents.
Build insight I didn't expect: keeping the agent in the SAME container as the user's repo is the unlock. Cursor/Lovable bolt an AI onto your editor; this bolts an AI directly onto your deployed app. The agent can git status, run your tests, hot-restart your server, all in one place. iOS becomes the chat surface, container does the work.
Hardest parts: Stripe webhook silent failures (metadata field-name drift), Fargate task rehydration across deploys, container disk quotas.
vibekit.bot for the demo. Feedback welcome on the wedge, the pricing, the arch.
609.sol retweeted
25+ free models now on VibeKit! Today's @OpenRouter integration unlocks the entire free pool.
More of VibeKit is free to use than ever — build, ship, iterate without a credit card.
