Today's trending repos tell a coherent story: the agentic development stack is accumulating its own layer of tooling. Developers aren't just using AI assistants anymore — they're designing reusable skills, plugin ecosystems, and security frameworks specifically for agent-driven workflows. The scaffolding around agentic AI is quietly becoming its own engineering discipline.
Matt Pocock's Real-World Agent Skills Go Viral
TypeScript educator Matt Pocock open-sourced mattpocock/skills, a collection of skills pulled straight from his personal .agents directory. The pitch is blunt: "Skills for Real Engineers" — practical, battle-tested agent instructions rather than contrived tutorial examples. The repo picked up over 3,300 new stars today alone. For developers shaping their own coding agent setups, it's a rare look into how a working practitioner actually structures the AI instructions they use day-to-day.
obra/superpowers: A Skills Framework With a Methodology
obra/superpowers positions itself as more than a skills collection — it's an opinionated framework with a software development methodology baked in. It gathered nearly 800 new stars today. The project treats agent skills as first-class software artifacts with their own structure and lifecycle, not loose prompt files scattered across a repo. As teams start sharing and maintaining skills across projects, frameworks like this are shaping how the discipline organizes itself.
Cursor Publishes Its Official Plugin Specification
cursor/plugins landed on trending with close to 400 new stars — significant because it's the official plugin spec from the Cursor team itself. It defines how third-party developers can build and distribute plugins for the AI code editor, something previously left to community convention. Formalizing a plugin spec signals that Cursor is betting on an ecosystem and gives extension authors a stable contract to build against.
Tencent Open-Sources an AI Infrastructure Red Teaming Platform
Tencent released AI-Infra-Guard, a full-stack platform for red-teaming AI infrastructure. It covers threat scanning across MCP servers, model endpoints, and associated data pipelines, mapping findings to frameworks like MITRE ATT&CK. With AI infrastructure expanding across organizations faster than security teams can audit it, tooling that specifically targets the agentic attack surface is long overdue. The structured output is designed to slot into existing security workflows rather than sit in isolation.
github.com/Tencent/AI-Infra-Guard
OpenViking: A Self-Evolving Context Store for Agents
ByteDance's Volcano Engine team released OpenViking, described as a "self-evolving context database" that unifies memory, RAG, and skills for AI agents in a single store. The core idea is that an agent shouldn't juggle separate memory backends, retrieval systems, and skill registries — a unified, append-only event log can serve all three. It's early traction with 148 new stars today, but the architectural direction reflects where serious agent infrastructure is heading.