Highlight of the Week
Karpathy’s Autoresearch
A self-contained repo for an autonomous LLM training loop. An AI agent iteratively improves training code and hyperparameters based on human prompts, autonomously driving validation loss improvements. This is the “AI improving AI” loop made concrete and reproducible.
Agentic Engineering & Strategy
Harness Engineering as Cybernetics
Harness engineering is the modern iteration of cybernetic control systems. The key insight: success relies not on agent capability, but on calibrating feedback loops with system-specific knowledge — architecture docs, linters, golden principles. This maps directly to how Claude Code skills work: the skill document is the feedback loop calibration.
Anthropic’s Official Skills Guide
Anthropic published “The Complete Guide to Building Skills for Claude” — the foundational patterns behind the tools and frameworks released this week. If you’re building Claude Code skills, this is the canonical reference.
“Everything is Context” — File-System-as-Memory
Research formalizing the file-system-as-memory architecture: memory, tools, and knowledge stored as versioned files (Markdown/JSON) for traceability, auditability, and persistence. This is exactly the pattern OpenClaw uses — and it’s now getting academic validation.
Engineering & Reverse Engineering
Codex 5.4 Reverse Engineers a DOS Game
Codex 5.4 autonomously reverse-engineered SkyRoads from a bare DOS executable. In 6 hours, the agent unpacked assets, disassembled the EXE, rebuilt the renderer, and recompiled a functional version in Rust. No human intervention.
office-agents SDK
Renamed open-excel to office-agents and published the SDK to NPM — a lightweight framework for building AI plugins for Office-like flows (Excel, PowerPoint, etc.).
SuperClaude Framework
Open-source framework for Claude Code with pre-built commands and agentic structures for automating coding workflows.
AI Services & Business
The AI Services Playbook
A comprehensive guide on identifying high-value “boring” niches (HVAC, law firms, accounting) and positioning AI as an employee amplifier rather than a replacement. The framing matters — “amplifier” sells where “replacement” triggers resistance.
Google’s PaperBanana
An agentic framework that automatically generates publication-ready academic illustrations from methodology descriptions and captions. No manual design tools needed.
Case Studies
The “Larry” Case Study — $7k/Month Autonomous Business
An agent called “Larry” running on a local gaming PC autonomously manages a business — marketing, development, analytics — generating over $7k/month. This led to the creation of LarryBrain, a marketplace for verified AI agent skills.
Biological DOOM Agent
800,000 biological human brain cells in a petri dish have been interfaced with DOOM, exhibiting “goal-directed learning” by adapting and improving at the game in response to failure. Not practical, but conceptually wild.
Creative & Community
AI Cinema Production
AI agents (Luma) significantly lower the barrier to producing high-end cinema adaptations — showcasing a Red Rising teaser that would have been impossible at indie budgets a year ago.
Evangelion UI Pack for OpenClaw
An Evangelion-inspired UI design skill pack for OpenClaw/Claude. Form meets function.
Themes This Week
| Theme | Signal |
|---|---|
| Feedback loops > raw capability | Harness engineering, skills guides, “Everything is Context” all point the same direction |
| Agents running businesses | Larry generating $7k/month autonomously is proof-of-concept becoming proof-of-revenue |
| Reverse engineering at scale | Codex 5.4 on SkyRoads shows agents can now handle truly adversarial, undocumented codebases |
| Academic validation of builder patterns | File-system-as-memory, agentic skills — things builders discovered empirically are now getting formal treatment |