Dev Digest — March 19, 2026

🔥 HOT RELEASES

JetBrains Air — Agentic Development Environment (Public Preview) JetBrains launches Air, a dedicated agentic dev environment built on 26 years of IDE experience. Unlike chat-window bolt-ons, Air is purpose-built for delegating tasks to multiple AI agents (Codex, Claude Agent, Gemini CLI, Junie) running concurrently. Supports Agent Client Protocol (ACP), Docker sandboxing, Git worktrees. Free for macOS, Windows/Linux coming soon. This is JetBrains' answer to Cursor — but with multi-agent orchestration as the core metaphor. 🔗 https://blog.jetbrains.com/air/2026/03/air-launches-as-public-preview-a-new-wave-of-dev-tooling-built-on-26-years-of-experience/

Java 26 Released (March 17) 10 JEPs including ahead-of-time object caching for any GC, lazy constants for performance, HTTP/3 support via QUIC, and several Project Valhalla foundations. First non-LTS since Java 25, focused on library improvements, language innovation, and security. 🔗 https://www.infoq.com/news/2026/03/java26-released/

Google AI Studio — Full-Stack Vibe Coding (Just Dropped) Google launched full-stack vibe coding in AI Studio powered by Gemini 3.1 Pro. The wild part: you can prompt-generate real-time multiplayer games (they demo'd a first-person laser tag game). Persistent sessions, resume across devices, deploy from prompt to production. 🔗 https://blog.google/innovation-and-ai/technology/developers-tools/full-stack-vibe-coding-google-ai-studio/

Apple M5 Pro & M5 Max MacBook Pro New Fusion Architecture (two dies, one SoC), 18-core CPU, up to 4x AI performance vs previous gen. Relevant for local LLM inference and on-device ML development. 🔗 https://www.apple.com/newsroom/2026/03/apple-introduces-macbook-pro-with-all-new-m5-pro-and-m5-max/

GPT-5.4 + Codex Updates OpenAI released GPT-5.4 (March 5) with native computer use, tool search (47% token reduction in tool-heavy workflows), and 1M context window in Codex mode. Latest Codex updates add full-resolution image inspection, realtime transcription mode, v2 filesystem RPCs with Python SDK, and guardian-based Smart Approvals. 🔗 https://releasebot.io/updates/openai

🧪 INTERESTING REPOS

karpathy/autoresearch — Autonomous AI Research Karpathy's 630-line Python script that ran 50 ML experiments overnight on one GPU without human input. Autonomous experiment loop for LLM training research. 30K+ stars in one week — one of the fastest-growing repos in GitHub history. 💻 https://github.com/karpathy/autoresearch

vectorize-io/hindsight — Learning Agent Memory AI agent memory system hitting 91% on LongMemEval. Multi-strategy retrieval (semantic, temporal, entity, keyword), fact extraction from conversations, MCP server support. A serious alternative to Mem0. 💻 https://github.com/vectorize-io/hindsight

GitNexus — Zero-Server Code Knowledge Graphs Client-side knowledge graph creator that runs entirely in your browser. Drop in a GitHub repo or ZIP, get an interactive knowledge graph with built-in Graph RAG agent. Great for onboarding AI agents onto unfamiliar codebases via MCP. 🔗 https://sinapti.ca/post/en/gitnexus-the-knowledge-graph-that-indexes-your-code-so-ai-ag-3dyocmpd

msitarzewski/agency-agents — Complete AI Agency Kit 51 specialized agent roles (frontend, marketing, fact-checking, etc.) in Shell scripts. +19,856 stars this week. Viral on X but HN is skeptical — more template than substance, but useful for quickly scaffolding multi-agent workflows. 💻 https://github.com/msitarzewski/agency-agents

BigBodyCobain/Shadowbroker — OSINT Tool Integrates 15 real-time data sources from corporate jet tracking to satellite orbits. Hit 304 HN points — the week's highest discussion score. TypeScript. 💻 https://github.com/BigBodyCobain/Shadowbroker

HKUDS/CLI-Anything — Universal CLI Agent Turn any task into a CLI command with AI. 2,707 stars in first week. Python. 💻 https://github.com/HKUDS/CLI-Anything

NousResearch/hermes-agent — Collaborative AI Agent NousResearch's agent designed for collaborative growth, built on Hermes models. 3,831 stars. 💻 https://github.com/NousResearch/hermes-agent

alibaba/OpenSandbox — Sandboxed Code Execution Alibaba's sandboxing solution for AI agent code execution. 7,407 stars and climbing. 💻 https://github.com/alibaba/OpenSandbox

🎥 WORTH WATCHING

Claude Code + Karpathy's Autoresearch = The New Meta How to combine Claude Code with autoresearch for autonomous ML experimentation. 🔗 https://www.youtube.com/watch?v=4Cb_l2LJAW8

METR Talk: "Reconciling Impressive AI Benchmark Performance with Limited Developer Productivity" Joel Becker (METR) presents data on why AI benchmarks don't translate to dev productivity gains yet. Essential viewing for anyone managing AI coding expectations. 🔗 https://www.youtube.com/watch?v=Sf9vENm2GnE

Small Language Models Are the Future: Fine-Tuning AI That Runs on Your iPhone Talk on SLMs and practical on-device inference. Relevant for mobile-first development. 🔗 https://www.youtube.com/watch?v=EXB8HokGVMI

AI Coding Is Making Devs Miserable (2026 Predictions) TheSeniorDev's take on the friction between AI coding hype and daily reality. 🔗 https://www.youtube.com/watch?v=8I_yeica3r4

💡 TECHNIQUES & IDEAS

The Developer-as-Manager Shift Business Insider reports Google and others are using AI to handle bulk coding, shifting developers toward decision-making and oversight roles. The NYT's long-form "Coding After Coders" piece describes devs who "rarely write code" anymore — spending days talking to AI agents in plain English and reviewing their plans. 🔗 https://www.businessinsider.com/ai-coding-changing-software-developer-role-2026-3 🔗 https://www.nytimes.com/2026/03/12/magazine/ai-coding-programming-jobs-claude-chatgpt.html

OpenAI Codex Smart Approvals Guardian-based approval system for autonomous coding agents. Instead of blanket permissions or constant interruption, the guardian evaluates each action against policies. This is the pattern for safe agent autonomy.

Multi-Agent Orchestration as the Default JetBrains Air, Codex's spawn_agent/wait_agent pattern, and the agency-agents repo all point the same direction: the future isn't one AI coding assistant, it's managing a team of them with different specializations.

🔮 EMERGING TRENDS

GlassWorm: Supply Chain Attack via Invisible Unicode 400+ compromised repos on GitHub, npm, and VSCode/OpenVSX. Uses invisible Unicode characters to hide malicious code that evades standard code review. Queries Solana blockchain every 5 seconds for C2 updates. Steals npm/GitHub/Git credentials and crypto assets. Check your dependencies. 🔗 https://www.bleepingcomputer.com/news/security/glassworm-malware-hits-400-plus-code-repos-on-github-npm-vscode-openvsx/

Agent Client Protocol (ACP) Gaining Traction JetBrains and Zed are pushing ACP as the standard for AI agent interoperability. The ACP Agent Registry launched in January, and Air's adoption signals it's becoming the plumbing layer for multi-agent IDEs.

Skills Ecosystem Exploding 5 of the top 10 new GitHub repos this week are agent Skills (SwiftUI, Medical, etc.). The pattern of packaging domain expertise as reusable agent skills is becoming a category unto itself. OpenClaw's skills ecosystem is part of this wave — the NYT and CNBC both covered OpenClaw's rapid GitHub adoption this week.

Vibe Coding Goes Full-Stack Google's AI Studio update shows vibe coding isn't just prototyping anymore — it's multiplayer games with real-time networking, persistent sessions, and production deployments. The gap between "prompt" and "product" keeps shrinking.

Note: X/Twitter API credits exhausted — this digest was compiled from web research. Quality may vary from X-sourced editions.

Full digest: https://digest.99.cooking