Coding agents & LLM mechanics

• Claude Code — Anthropic's terminal coding agent repo, still the category's reference implementation.
• andrej-karpathy-skills — single CLAUDE.md distilling Karpathy's observation that models run with wrong assumptions instead of asking, packaged as a drop-in guardrail.
• The Evolution of the Agent Harness — argues agents "started working" around Christmas 2025 from model capability crossing a threshold plus harnesses finally maturing around it.
• New MCP Roadmap — maintainers publish next-spec direction; community pushback says MCP's bespoke protocol was a mistake and stateless HTTP is the real fix.
• A week of using Codex more than Claude — dev finds Codex faster and cleaner-coded, flags Opus 5.0 as a regression vs 4.8.
• Zero — Vercel's new programming language built specifically for AI agents.
• Munder Difflin — free multi-agent harness wrapping your existing CLI/subscription to run an "office of clones" locally; community split on practicality vs novelty.
• Autolith — terminal coding agent with a live Common Lisp runtime for recursive inference over large corpora; thin Lisp training data limits real-world uptake.
• nodeterm — tmux-backed canvas turning terminals and parallel Claude Code sessions into a draggable Trello-style board.
• Claude's output is not optimised for humans — thread arguing Claude increasingly formats for downstream agent consumption, not human readers.
• Why your local LLM feels dumber than it is — silent GGUF metadata loss breaks chat templates and tanks perceived quality, not quantization itself.
• More than just code review — Simon Willison on verifying agent-written code without eyeballing every line, the real skill bottleneck.
• openai/codex — OpenAI's terminal coding agent, direct Claude Code competitor.
• Agents Never Sleep — launch for agents that keep running with the laptop lid closed.
• NanoGPT Speedrun Frontier — 153 autonomous runs across 18 frontier models racing the nanoGPT speedrun; Fable 5 leads on validated closure, but inconsistent effort settings undercut cross-model comparison.
• 10% worse, 100x cheaper, 10000x faster: Why Simulation is taking over — argues cheap-fast-good-enough simulation is winning, ties together GLM/Poolside/AI-for-Science threads.
• Claude workflow for Android execution — community writeup wiring Claude into an Android execution loop.
• Quoting Linus Torvalds — Torvalds credits AI as a "tireless helper" in a debug-session-from-hell, even after it insisted the bug was unsolvable.

CLI/TUI

• llm 0.33 — Simon Willison's llm CLI upgrades to OpenAI Python 3.x, swaps httpx for httpx2.
• simonw/llm — the CLI/library underlying it, run any LLM (Claude, GPT, Gemini, local models) from the terminal.

Vibe-coding & repos

• Post-quantum ciphertext diary entirely in URLs — vibe-coded diary distributing PQ-encrypted entries as shareable URLs, no backend.
• One app to replace Illustrator, Lightroom, and most of After Effects — solo-built creative suite replacement, Figma-equivalent planned next.

Agentic SaaS

• n8n — fair-code workflow automation with native AI agent building, 1500+ integrations, self-host or cloud.
• sub2api — OSS relay unifying Claude/OpenAI/Gemini/Grok subscriptions behind one API for cost-sharing; project itself flags likely ToS violation with upstream providers.
• Open Analytics — AI-native Google Analytics alternative.