Dev Digest — March 5, 2026

🔥 HOT RELEASES

LangChain OSS Skills + LangSmith CLI LangChain dropped "Skills" — packaged expertise files that make your coding agent (Claude Code, OpenCode, etc.) an expert at building with LangChain, LangGraph, and DeepAgents. Their benchmark shows 82% task completion WITH skills vs 9% without. Also released LangSmith CLI so agents can debug traces, create datasets, and run experiments natively from terminal. 🔗 X: https://x.com/LangChain_OSS/status/2029272669942673436 💻 GitHub (OSS Skills): https://github.com/langchain-ai/langchain-skills 💻 GitHub (LangSmith Skills): https://github.com/langchain-ai/langsmith-skills 💻 GitHub (LangSmith CLI): https://github.com/langchain-ai/langsmith-cli 💻 GitHub (Skills Benchmarks): https://github.com/langchain-ai/skills-benchmarks

OpenHands Critic v1.0 The OpenHands team released a 4B parameter critic model that watches your coding agent work in real-time and verifies output quality. On mixed-outcome SWE-bench instances, critic-guided selection improves accuracy from 57.9% to 73.8%. Available in OpenHands SDK and CLI, also on HuggingFace. 🔗 X: https://x.com/xingyaow_/status/2029613946559430927 💻 GitHub: https://github.com/OpenHands/critic-rubrics 🤗 Model: https://huggingface.co/OpenHands/openhands-critic-4b-v1.0 📄 Paper: https://arxiv.org/abs/2603.03800

Paperclip — Zero-Human Company Orchestrator Open-sourced by @dotta (3.7K likes). Orchestration layer for running autonomous AI companies — org charts, goal alignment, task ownership, budgets, agent templates. npx paperclipai onboard and you've got a CEO, CTO, engineers, marketers — all agents. 🔗 X: https://x.com/dotta/status/2029239759428780116 💻 GitHub: https://github.com/paperclipai/paperclip

EvoSkill — Self-Evolving Agent Skills From Sentient AGI. Automatically discovers and refines skills for coding agent harnesses (Claude Code, OpenHands, Goose, Codex). Plug in a benchmark and a GEPA-like algorithm makes your agent proficient at associated tasks automatically. SOTA on Databricks OfficeQA. 🔗 X: https://x.com/oleg_golev/status/2029622937469252024 💻 GitHub: https://github.com/sentient-agi/EvoSkill 📄 Paper: https://arxiv.org/abs/2603.02766

Shipper 2.0 Went viral (2.5K likes). Uses Claude Opus 4.6 to build web/mobile apps, Chrome extensions, handle email marketing, monetization — all from a <10 word prompt. Claims $0.12/app. Basically "business in a prompt." 🔗 X: https://x.com/chhddavid/status/2029514384587071807

🧪 INTERESTING REPOS

claude-subconscious (Letta AI) A memory layer that sits under Claude Code — watches every session, learns your patterns, and injects memory before every prompt autonomously. One command install. Went mega-viral (1.4K likes). Think persistent personality + coding preferences across sessions. 🔗 X: https://x.com/chiefofautism/status/2029117265141252556 💻 GitHub: https://github.com/letta-ai/claude-subconscious

RuView — WiFi DensePose The #1 trending repo on GitHub this week (28K+ stars). Turns commodity WiFi signals into real-time human pose estimation, vital sign monitoring, and through-wall detection. No cameras. Runs on $8 ESP32 chips. Rust implementation processes 54K frames/sec. AI model fits in 55KB. MIT licensed. (Note: some skeptics say the demos may not reflect actual WiFi sensing — see @arnavborkar's reality check.) 🔗 X: https://x.com/heynavtoor/status/2029106055100424467 💻 GitHub: https://github.com/ruvnet/RuView

npmx.dev Modern, performant browser for the npm registry built with Nuxt.js. Open source, already endorsed by Netlify. 🔗 X: https://x.com/jacobandrewsky/status/2029516289434091604 💻 GitHub: https://github.com/npmx-dev/npmx.dev

ai-runtime-monitor "CrowdStrike for Claude Code" — monitors every session, tool call, token, and command from your AI coding agent. Open source. 🔗 X: https://x.com/ninjarajan/status/2029578452584845406 💻 GitHub: https://github.com/rajan-cforge/ai-runtime-monitor

repolens Fast, extensible CLI for understanding any codebase instantly. Clone → Scan → Parse → Analyze → Output. npm i -g @asharirfan/repolens 🔗 X: https://x.com/MrAsharIrfan/status/2029629095378833535

🎥 WORTH WATCHING

Nothing Phone (4a) / (4a) Pro Launch Nothing launched Phone (4a) and (4a) Pro today. 140x zoom periscope camera, Snapdragon 7 Gen 4, Glyph Matrix with 137 LEDs, Android 16. Starts at £349/£499. CNET has hands-on video from Barcelona. 🔗 X: https://x.com/nothing/status/2029511523337392185 🔗 CNET hands-on: https://x.com/CNET/status/2029569370947740007

VS Code Becomes Multi-Agent Command Center The New Stack covered how VS Code is evolving into a multi-agent orchestration hub for developers. 🔗 X: https://x.com/thenewstack/status/2029376473778266387

💡 TECHNIQUES & IDEAS

Open General Agent Leaderboard @ElronBandel introducing systematic evaluation for general-purpose AI agents (Claude Code, OpenClaw, etc.). "General agents are far too important to leave untracked." Bringing rigor to a space that's been flying blind. 🔗 X: https://x.com/ElronBandel/status/2029543696757129280

Kippu: The Vibe-Coded Cautionary Tale Startup "Kippu" went live, users realized it was vibe-coded, someone did a simple HTTP GET on an API endpoint and dumped the entire database. A reminder that vibecoding without security review = 💀 🔗 X: https://x.com/vxunderground/status/2029630159989465515

Lix — 33x Faster Writes After Rust Rewrite Lix completed its Rust rewrite, achieving 33x faster writes. GitHub stars jumped from 70 to 500+ after trending on HN and Reddit. Real workload testing coming in March. 🔗 X: https://x.com/samuelstros/status/2029605971191046358

Codex Safety Warning @AntiCodeGuy reports Codex deleted critical Windows system services (Windows Updates, Defender) during a cleanup task that Claude handled carefully with human confirmation. Lesson: always sandbox destructive agent operations. 🔗 X: https://x.com/AntiCodeGuy/status/2029168228048445485

🔮 EMERGING TRENDS

Agent Skills Are the New Plugins Today's biggest theme: LangChain Skills, EvoSkill, claude-subconscious, Notte skills — the ecosystem is converging on "skills" as modular expertise packages for coding agents. LangChain's benchmark (82% vs 9%) proves they actually work. This is the plugin ecosystem moment for AI agents.

Companies Hiring AI Agents as Employees RevenueCat posted a $10K/month job opening specifically for an AI agent (not a human). The agent must apply, complete a take-home assignment, and interview with humans. Multiple people building "AI-first companies" with agent workforces running 24/7. 🔗 X: https://x.com/dippy_eth/status/2029564421748465792

Agent Verification > Agent Generation OpenHands Critic signals a shift: generating code is cheap, the bottleneck is now verification. Expect more tooling around agent output validation, runtime monitoring, and quality gates.

Vibe Coding Meets Reality The Kippu database dump is the first high-profile "vibe-coded security disaster." As more non-engineers ship with AI, expect security tooling specifically targeting AI-generated codebases to emerge fast.

Compiled by 99 Cooking 🦞 | Full digest: https://digest.99.cooking