Dev Digest — March 17, 2026

🔥 HOT RELEASES

Java 26 (JDK 26) Oracle dropped Java 26 today at JavaOne — 10 JEPs focused on AI readiness, performance (G1 GC improvements), crypto modernization, and language cleanup. Finally removes the Applet API. Also announced the "Java Verified Portfolio" — a curated, supported stack for enterprise devs. 🔗 https://www.infoworld.com/article/4145086/oracle-unveils-the-java-verified-portfolio.html

VS Code 1.111 — First Weekly Release After 10 years of monthly releases, VS Code has gone weekly. v1.111 introduces "Autopilot mode" — AI agents that can run tasks without manual approval. PMs (not just engineers) can now define and prototype features using AI agents. Credits GitHub Copilot for making this cadence possible. 🔗 https://www.theregister.com/2026/03/11/visual_studio_code_moves_to/

Superpowers — Skills Framework for Coding Agents (40.9K stars) By Jesse Vincent (obra). Not another agent framework — it's a structured methodology layer on TOP of Claude Code/Codex. Composable skills library of proven techniques, patterns, and tools. Gained 1,867 stars in a single day. 40K+ devs saying "agents need better structure." 💻 https://github.com/obra/superpowers

🧪 INTERESTING REPOS

karpathy/autoresearch — 22,983 stars in 3 days Karpathy's autonomous ML research lab. A 630-line Python script (MIT license) that lets AI agents iterate on code experiments overnight. Point it at a problem, go to sleep, wake up to results. The pattern is already being generalized beyond ML. 💻 https://github.com/karpathy/autoresearch

vectorize-io/hindsight — Agent Memory That Learns Multi-strategy memory retrieval for AI agents — semantic similarity, temporal proximity, entity relationships, keyword matching. Extracts atomic facts from conversations. Works with local models via Ollama. Direct competitor to Mem0 but with deeper retrieval. 💻 https://github.com/vectorize-io/hindsight

HKUDS/CLI-Anything — 12.1K stars Turn any software into a CLI tool. Analyzes source code/SDK, then auto-generates runnable CLI commands. From HKU Data Science lab. Growing fast with the "agents need tools" wave. 💻 https://github.com/HKUDS/CLI-Anything

abhigyanpatwari/GitNexus — Zero-Server Code Intelligence Client-side knowledge graph for codebases. Drop in a GitHub repo → get an interactive knowledge graph with a Graph RAG agent. Runs entirely in-browser (Tree-sitter WASM). Also has CLI + MCP integration so coding agents get architectural awareness before making edits. 💻 https://github.com/abhigyanpatwari/GitNexus

NousResearch/hermes-agent — 3,831 stars Self-improving AI agent that writes new functions for itself based on repeated interaction patterns. Persistent memory + skill acquisition. "An agent that grows with you." Works with OpenAI API or local models. 💻 https://github.com/NousResearch/hermes-agent

BigBodyCobain/Shadowbroker — 1,415 stars Real-time OSINT dashboard aggregating 15 data sources: corporate jet tracking, spy satellite orbits, ship tracking, seismic events, conflict zones. Hit 304 HN points. Self-hosted "God-Eye" for the cyberpunk-curious. 💻 https://github.com/BigBodyCobain/Shadowbroker

msitarzewski/agency-agents — 25,639 stars (+19.8K/week) Pre-defined AI agent roles for a complete "AI agency" — frontend engineer, marketing, social media, fact-checker. Each agent has personality, workflow, and deliverable format. Shell-based, connects to any LLM backend. 💻 https://github.com/msitarzewski/agency-agents

🎥 WORTH WATCHING

Karpathy's "autoresearch" explained in 24 minutes Clear breakdown of how autoresearch works and why it matters beyond ML. 🔗 https://www.youtube.com/watch?v=qb90PPbAWz4

Claude Code + Autoresearch = SELF-IMPROVING AI Practical walkthrough of combining Claude Code with the autoresearch pattern. 🔗 https://www.youtube.com/watch?v=4Cb_l2LJAW8

"Coding After Coders" — NYT Magazine deep dive Clive Thompson spoke to 70+ devs at Google, Amazon, Microsoft, Apple about how AI is transforming programming. Features Claude Code multi-agent workflows in action. The most substantial mainstream piece on AI coding yet. 🔗 https://www.nytimes.com/2026/03/12/magazine/ai-coding-programming-jobs-claude-chatgpt.html

💡 TECHNIQUES & IDEAS

Autoresearch as a General Pattern Karpathy built it for ML, but the community is already applying the iterate-autonomously pattern to web dev, trading bots, and infrastructure optimization. Someone built a Claude Code skill that generalizes it to any task — 231 upvotes on r/ClaudeCode. 🔗 https://www.reddit.com/r/ClaudeCode/comments/1rsur5s/

Conformance-Driven Development From the Pragmatic Summit fireside: instead of TDD, define "conformance specs" that AI agents must satisfy. Red-green but for agents. Also discussed: prompt-injection risks in agentic workflows and "Showboat" manual testing for agent outputs. 🔗 https://letsdatascience.com/news/developers-adopt-agentic-tools-for-production-coding-b82b6ec0

PMs Prototyping with AI Agents VS Code's new workflow lets Product Managers create test plans from feature request issues with one click, then AI agents prototype the feature. Engineers review, not build from scratch. The role boundary is shifting.

🔮 EMERGING TRENDS

The Skills Explosion 5 of the top 10 new GitHub repos this week are "skills" — structured instruction sets for AI coding agents. Superpowers (40K stars), claude-skills, SwiftUI-Agent-Skill, Medical Skills, agency-agents. The meta-game is now about what you tell the agent, not which agent you use.

Agent Memory Gets Serious Hindsight (vectorize-io), Hermes Agent (NousResearch), and Mem0 are all competing on how agents remember across sessions. Multi-strategy retrieval, fact extraction, self-improving skills. Memory is becoming the moat.

Supply Chain Attacks Scale with AI Glassworm campaign compromised 151+ GitHub repos using invisible Unicode characters, plus npm packages and VS Code extensions. Uses Solana blockchain for C2. Researchers believe attackers are using LLMs to generate convincing malicious code at scale. Worth auditing your deps. 🔗 https://arstechnica.com/security/2026/03/supply-chain-attack-using-invisible-code-hits-github-and-other-repositories/

91% of Software Companies Now Use AI for Development GoodFirms survey (today): 91% of software companies use AI to cut development costs. Stack Overflow puts it at 84% of individual devs. We've crossed from "early adopters" to "everyone."

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