📬 AI Daily Digest · 2026-05-21
🔬 New AI Craft
1. Formal Verification Gates for AI Coding Loops
Embedding formal verification as a quality gate in AI coding loops — code must pass mathematical correctness proof before merge. Unlike the Agent Plan workflow (which relies on human review), this provides mathematical-level guarantees that catch logic bugs and edge cases. Challenges remain around automation and performance overhead of formal verification tools.
https://news.ycombinator.com/item?id=48205676
2. ProofShot — Visual Verification for AI Coding Agents
A tool that lets AI agents take screenshots and compare UI changes to automatically verify frontend output correctness. Think visual regression testing for agents. Complements AI-generated unit tests (CodiumAI/Diffblue) — unit tests verify logic, ProofShot verifies rendered output.
https://news.ycombinator.com/item?id=48205676
3. AGENTS.md — Open Standard for Guiding Coding Agents
A new open format using AGENTS.md files to describe repo structure, build commands, test strategies — everything an AI agent needs to know. Think CLAUDE.md but as a universal protocol. Multiple agent tools (OpenCode, Claude Code) are adding support, moving agent collaboration from "hint at it in README" to "follow a standard contract."
https://news.ycombinator.com/item?id=48205676
🛠️ Tools & Tips
1. obra/superpowers — 200K ⭐ Agentic Skills Framework & Methodology
A full agentic skills framework and software development methodology that breaks down development into executable skill units for AI agents. 200K stars signals strong community validation.
https://github.com/obra/superpowers
2. andrej-karpathy-skills — 140K ⭐ Claude Code Enhancement from Karpathy
A single CLAUDE.md file distilling Andrej Karpathy's observations on LLM coding pitfalls. Drop it in and Claude Code behavior improves immediately — no installation needed.
https://github.com/multica-ai/andrej-karpathy-skills
3. agentmemory — #1 Persistent Memory for AI Coding Agents
Benchmark-validated persistent memory that lets AI coding agents remember project context, user preferences, and past decisions across sessions.
https://github.com/rohitg00/agentmemory
⭐ Open Source Highlights
1. OpenCode — Open source AI coding agent (1274 HN points)
Terminal-native open source AI coding agent, going head-to-head with Claude Code.
https://github.com/colbymchenry/opencode
2. oh-my-pi (⌥) — AI Coding Agent for the Terminal
Hash-anchored edits, LSP support, sub-agent architecture, zero-config setup. 5.3K stars in days.
https://github.com/can1357/oh-my-pi
3. Qwen3.6-35B-A3B — Open Agentic Coding Model from Alibaba
35B params with only 3B active (MoE), optimized for agentic coding. 1274 HN points.
https://qwen.ai/blog?id=qwen3.7
📰 Industry News
- GitHub confirms breach of 3,800 repos via malicious VSCode extension
- Google Chrome silently installs 4GB AI model, sparking privacy backlash
- Anthropic expanding to Colossus2 datacenter, will use NVIDIA GB200
🚀 Major Releases
- Qwen3.7-Max: The Agent Frontier — Alibaba's agent-optimized model
- Anthropic releases Claude Sonnet 4.5, claiming top coding performance
- OpenAI reportedly preparing GPT-5 launch in August
