Autonomous Agent Loops Are Real: Karpathy, Stripe, and Anthropic's Dreaming
Three verified cases of AI agents running in loops: Karpathy's 700 experiments in two days, Stripe's 1,300 weekly pull requests, Anthropic's dreaming.
Build journal
Software problems, and how I solved them: architecture, tools, costs, accessibility. For people who build.
28 articles · topic: All · sort: newest first
Three verified cases of AI agents running in loops: Karpathy's 700 experiments in two days, Stripe's 1,300 weekly pull requests, Anthropic's dreaming.
claude-video adds a /watch command to Claude Code: paste a YouTube, Loom or TikTok URL and the agent sees frames and reads the transcript. Costs and limits.
Karpathy popularized letting an agent write your personal wiki instead of maintaining it yourself. The real structure, the real setup time, and what changes.
A viral list of 22 Claude Code skills, checked one by one: all real and active, except one number borrowing another project's fame. How to install them.
ECC (Everything Claude Code) promises ready-to-use agents, skills, commands and a security scanner. What the repo really contains, and what to check first.
The google/skills repo ships ready-to-use Google Cloud skills, compatible with Claude Code and other agents: over 100 now, up from 13 at launch in March 2026.
claude-howto is a free, open source Claude Code guide, beginner to advanced: slash commands, memory, skills, subagents, MCP, hooks. 10 modules, 11-13 hours.
How kimi-k3-in-c works: 4-bit quantization, a resident dense trunk, disk-streamed MoE experts and an LRU cache. The real bottleneck is disk, not compute.
Tech story: shipping a multilingual, realtime tournament platform with complex bracket formats (Swiss, double elimination, pools, etc.) for GIGA'GAMES Esports.