todai
Wednesday, 20 May
5 min read · 839 words
  • humanlayer/12-factor-agents — 733★ today (21,098★ total) — 12 opinionated rules for building LLM-powered software that's actually production-stable — think "12 Factor App" applied to agent design | https://github.com/humanlayer/12-factor-agents

TODAY'S ITEMS

1. Andrej Karpathy Joins Anthropic's Pre-Training Team

  • Andrej Karpathy — OpenAI co-founder, former Tesla AI lead, and the person who coined "vibe coding" — has joined Anthropic's pre-training team, announced this morning.
  • He stated he's excited to return to R&D and plans to resume his education work "in time."
  • Source: @karpathy
  • Why it matters: Karpathy coined "vibe coding" and has shaped how the field thinks about AI-assisted development — his move to Anthropic pre-training is the clearest signal yet that Claude's coding capabilities are a top strategic priority.
  • Verified

2. MCP Tunnels + Self-Hosted Sandboxes Launch for Claude Managed Agents

  • Announced today at Code with Claude London: MCP tunnels (research preview) let Claude agents reach MCP servers inside your private network without exposing them to the internet; self-hosted sandboxes (public beta) move tool execution off Anthropic infrastructure and onto yours.
  • Both available now under the standard managed-agents-2026-04-01 beta header.
  • Source: Anthropic Platform Changelog
  • Why it matters: If you're building agents that need to reach internal databases, private APIs, or on-prem services, MCP tunnels remove the last reason to keep Claude agents off those systems.
  • Verified

3. Claude Code v2.1.144 — Background Sessions in /resume, Startup Hang Fixed

  • Background sessions (started via claude --bg or agent view) now appear in /resume alongside interactive ones; the up-to-75-second startup hang when api.anthropic.com is unreachable (captive portals, firewalls, VPNs) is fixed — side-channel API calls now time out after 15s.
  • Also: /model is now scoped to the current session only; press d in the model picker to set a new default for future sessions.
  • Source: Claude Code Releases
  • Why it matters: Update now if you run Claude Code behind a corporate firewall or VPN — the 75-second startup hang is fixed in this release.
  • Verified

4. Anthropic Launches Official Curated Claude Code Plugin Directory

  • anthropics/claude-plugins-official is live on GitHub — an Anthropic-managed directory of vetted plugins including internal Anthropic-built plugins and approved third-party ones with quality and security review.
  • Browse in Claude Code via /plugin > Discover, or install directly with /plugin install {name}@claude-plugins-official.
  • Source: GitHub
  • Why it matters: Instead of hunting community repos hoping they're reliable, you now have a curated directory maintained by Anthropic — every plugin in it passed a security and quality bar.
  • Verified

YOUR STACK — UPDATES

NEW TOOL / PRODUCT SPOTLIGHT

  • rtk (Rust Token Killer) — CLI proxy that compresses shell command outputs (git, grep, cat, test runs) before they hit your LLM context, cutting token usage ~80% on typical dev sessions. Single Rust binary, zero dependencies. brew install rtk then rtk init -g to auto-hook Bash so all AI tools get compressed output automatically | https://github.com/rtk-ai/rtk

PROMPT OF THE DAY

# Drop this into your CLAUDE.md — 4 principles from Karpathy's LLM
# coding failure modes (timed for today's announcement)

## Think Before Coding
State assumptions explicitly. When ambiguous, present options and ask
— don't pick silently and run. Push back if a simpler approach exists.
Stop when confused; name what's unclear.

## Simplicity First
Minimum code that solves the problem. No unrequested features, no
abstractions for single-use code, no speculative error handling.
If 200 lines could be 50, rewrite it.

## Surgical Changes
Touch only what you must. Don't improve adjacent code or formatting.
Match existing style. Every changed line must trace directly to the
user's request.

## Goal-Driven Execution
Transform requests into verifiable goals:
"Add validation" → "Write failing tests for invalid inputs, make them pass."
State a brief plan before starting multi-step tasks.

Drop this into your project's CLAUDE.md to directly counter the 4 failure modes Karpathy documented in LLM coding agents. Community source: https://github.com/multica-ai/andrej-karpathy-skills

LANDSCAPE NOTES